Merge branch 'master' into DPM++SDE

README.md

@@ -84,26 +84,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web
 - API
 - Support for dedicated [inpainting model](https://github.com/runwayml/stable-diffusion#inpainting-with-stable-diffusion) by RunwayML. 
 - via extension: [Aesthetic Gradients](https://github.com/AUTOMATIC1111/stable-diffusion-webui-aesthetic-gradients), a way to generate images with a specific aesthetic by using clip images embds (implementation of [https://github.com/vicgalle/stable-diffusion-aesthetic-gradients](https://github.com/vicgalle/stable-diffusion-aesthetic-gradients))
-
-## Where are Aesthetic Gradients?!?!
-Aesthetic Gradients are now an extension. You can install it using git:
-
-```commandline
-git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui-aesthetic-gradients extensions/aesthetic-gradients
-```
-
-After running this command, make sure that you have `aesthetic-gradients` dir in webui's `extensions` directory and restart
-the UI. The interface for Aesthetic Gradients should appear exactly the same as it was.
-
-## Where is History/Image browser?!?!
-Image browser is now an extension. You can install it using git:
-
-```commandline
-git clone https://github.com/yfszzx/stable-diffusion-webui-images-browser extensions/images-browser
-```
-
-After running this command, make sure that you have `images-browser` dir in webui's `extensions` directory and restart
-the UI. The interface for Image browser should appear exactly the same as it was.
+- [Stable Diffusion 2.0](https://github.com/Stability-AI/stablediffusion) support - see [wiki](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#stable-diffusion-20) for instructions
 
 ## Installation and Running
 Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for both [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended) and [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs.

launch.py

@@ -134,18 +134,19 @@ def prepare_enviroment():
 
     gfpgan_package = os.environ.get('GFPGAN_PACKAGE', "git+https://github.com/TencentARC/GFPGAN.git@8d2447a2d918f8eba5a4a01463fd48e45126a379")
     clip_package = os.environ.get('CLIP_PACKAGE', "git+https://github.com/openai/CLIP.git@d50d76daa670286dd6cacf3bcd80b5e4823fc8e1")
+    openclip_package = os.environ.get('OPENCLIP_PACKAGE', "git+https://github.com/mlfoundations/open_clip.git@bb6e834e9c70d9c27d0dc3ecedeebeaeb1ffad6b")
 
     xformers_windows_package = os.environ.get('XFORMERS_WINDOWS_PACKAGE', 'https://github.com/C43H66N12O12S2/stable-diffusion-webui/releases/download/f/xformers-0.0.14.dev0-cp310-cp310-win_amd64.whl')
 
-    stable_diffusion_repo = os.environ.get('STABLE_DIFFUSION_REPO', "https://github.com/CompVis/stable-diffusion.git")
+    stable_diffusion_repo = os.environ.get('STABLE_DIFFUSION_REPO', "https://github.com/Stability-AI/stablediffusion.git")
     taming_transformers_repo = os.environ.get('TAMING_TRANSFORMERS_REPO', "https://github.com/CompVis/taming-transformers.git")
     k_diffusion_repo = os.environ.get('K_DIFFUSION_REPO', 'https://github.com/crowsonkb/k-diffusion.git')
     codeformer_repo = os.environ.get('CODEFORMER_REPO', 'https://github.com/sczhou/CodeFormer.git')
     blip_repo = os.environ.get('BLIP_REPO', 'https://github.com/salesforce/BLIP.git')
 
-    stable_diffusion_commit_hash = os.environ.get('STABLE_DIFFUSION_COMMIT_HASH', "69ae4b35e0a0f6ee1af8bb9a5d0016ccb27e36dc")
+    stable_diffusion_commit_hash = os.environ.get('STABLE_DIFFUSION_COMMIT_HASH', "47b6b607fdd31875c9279cd2f4f16b92e4ea958e")
     taming_transformers_commit_hash = os.environ.get('TAMING_TRANSFORMERS_COMMIT_HASH', "24268930bf1dce879235a7fddd0b2355b84d7ea6")
-    k_diffusion_commit_hash = os.environ.get('K_DIFFUSION_COMMIT_HASH', "b325595b8a776d483f6935dfa7b45f01c27039e4")
+    k_diffusion_commit_hash = os.environ.get('K_DIFFUSION_COMMIT_HASH', "5b3af030dd83e0297272d861c19477735d0317ec")
     codeformer_commit_hash = os.environ.get('CODEFORMER_COMMIT_HASH', "c5b4593074ba6214284d6acd5f1719b6c5d739af")
     blip_commit_hash = os.environ.get('BLIP_COMMIT_HASH', "48211a1594f1321b00f14c9f7a5b4813144b2fb9")
 
@@ -179,6 +180,9 @@ def prepare_enviroment():
     if not is_installed("clip"):
         run_pip(f"install {clip_package}", "clip")
 
+    if not is_installed("open_clip"):
+        run_pip(f"install {openclip_package}", "open_clip")
+
     if (not is_installed("xformers") or reinstall_xformers) and xformers:
         if platform.system() == "Windows":
             if platform.python_version().startswith("3.10"):
@@ -196,7 +200,7 @@ def prepare_enviroment():
 
     os.makedirs(dir_repos, exist_ok=True)
 
-    git_clone(stable_diffusion_repo, repo_dir('stable-diffusion'), "Stable Diffusion", stable_diffusion_commit_hash)
+    git_clone(stable_diffusion_repo, repo_dir('stable-diffusion-stability-ai'), "Stable Diffusion", stable_diffusion_commit_hash)
     git_clone(taming_transformers_repo, repo_dir('taming-transformers'), "Taming Transformers", taming_transformers_commit_hash)
     git_clone(k_diffusion_repo, repo_dir('k-diffusion'), "K-diffusion", k_diffusion_commit_hash)
     git_clone(codeformer_repo, repo_dir('CodeFormer'), "CodeFormer", codeformer_commit_hash)

modules/hypernetworks/hypernetwork.py

@@ -38,7 +38,7 @@ class HypernetworkModule(torch.nn.Module):
     activation_dict.update({cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(torch.nn.modules.activation) if inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'})
 
     def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, weight_init='Normal',
-                 add_layer_norm=False, use_dropout=False, activate_output=False, last_layer_dropout=True):
+                 add_layer_norm=False, use_dropout=False, activate_output=False, last_layer_dropout=False):
         super().__init__()
 
         assert layer_structure is not None, "layer_structure must not be None"
@@ -154,16 +154,28 @@ class Hypernetwork:
                 HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init,
                                    self.add_layer_norm, self.use_dropout, self.activate_output, last_layer_dropout=self.last_layer_dropout),
             )
+        self.eval_mode()
 
     def weights(self):
         res = []
+        for k, layers in self.layers.items():
+            for layer in layers:
+                res += layer.parameters()
+        return res
 
+    def train_mode(self):
         for k, layers in self.layers.items():
             for layer in layers:
                 layer.train()
-                res += layer.trainables()
+                for param in layer.parameters():
+                    param.requires_grad = True
 
-        return res
+    def eval_mode(self):
+        for k, layers in self.layers.items():
+            for layer in layers:
+                layer.eval()
+                for param in layer.parameters():
+                    param.requires_grad = False
 
     def save(self, filename):
         state_dict = {}
@@ -367,13 +379,13 @@ def report_statistics(loss_info:dict):
 
 
 
-def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
+def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
     # images allows training previews to have infotext. Importing it at the top causes a circular import problem.
     from modules import images
 
     save_hypernetwork_every = save_hypernetwork_every or 0
     create_image_every = create_image_every or 0
-    textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, data_root, template_file, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork")
+    textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork")
 
     path = shared.hypernetworks.get(hypernetwork_name, None)
     shared.loaded_hypernetwork = Hypernetwork()
@@ -403,32 +415,30 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
     hypernetwork = shared.loaded_hypernetwork
     checkpoint = sd_models.select_checkpoint()
 
-    ititial_step = hypernetwork.step or 0
-    if ititial_step >= steps:
+    initial_step = hypernetwork.step or 0
+    if initial_step >= steps:
         shared.state.textinfo = f"Model has already been trained beyond specified max steps"
         return hypernetwork, filename
 
-    scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
-    
+    scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
+
     # dataset loading may take a while, so input validations and early returns should be done before this
     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
-    with torch.autocast("cuda"):
-        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size)
+
+    pin_memory = shared.opts.pin_memory
+
+    ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method)
+    
+    latent_sampling_method = ds.latent_sampling_method
+
+    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, latent_sampling_method=latent_sampling_method, batch_size=ds.batch_size, pin_memory=pin_memory)
 
     if unload:
         shared.sd_model.cond_stage_model.to(devices.cpu)
         shared.sd_model.first_stage_model.to(devices.cpu)
-
-    size = len(ds.indexes)
-    loss_dict = defaultdict(lambda : deque(maxlen = 1024))
-    losses = torch.zeros((size,))
-    previous_mean_losses = [0]
-    previous_mean_loss = 0
-    print("Mean loss of {} elements".format(size))
     
     weights = hypernetwork.weights()
-    for weight in weights:
-        weight.requires_grad = True
+    hypernetwork.train_mode()
 
     # Here we use optimizer from saved HN, or we can specify as UI option.
     if hypernetwork.optimizer_name in optimizer_dict:
@@ -446,131 +456,156 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
             print("Cannot resume from saved optimizer!")
             print(e)
 
+    scaler = torch.cuda.amp.GradScaler()
+    
+    batch_size = ds.batch_size
+    gradient_step = ds.gradient_step
+    # n steps = batch_size * gradient_step * n image processed
+    steps_per_epoch = len(ds) // batch_size // gradient_step
+    max_steps_per_epoch = len(ds) // batch_size - (len(ds) // batch_size) % gradient_step
+    loss_step = 0
+    _loss_step = 0 #internal
+    # size = len(ds.indexes)
+    # loss_dict = defaultdict(lambda : deque(maxlen = 1024))
+    # losses = torch.zeros((size,))
+    # previous_mean_losses = [0]
+    # previous_mean_loss = 0
+    # print("Mean loss of {} elements".format(size))
+
     steps_without_grad = 0
 
     last_saved_file = "<none>"
     last_saved_image = "<none>"
     forced_filename = "<none>"
 
-    pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step)
-    for i, entries in pbar:
-        hypernetwork.step = i + ititial_step
-        if len(loss_dict) > 0:
-            previous_mean_losses = [i[-1] for i in loss_dict.values()]
-            previous_mean_loss = mean(previous_mean_losses)
-            
-        scheduler.apply(optimizer, hypernetwork.step)
-        if scheduler.finished:
-            break
+    pbar = tqdm.tqdm(total=steps - initial_step)
+    try:
+        for i in range((steps-initial_step) * gradient_step):
+            if scheduler.finished:
+                break
+            if shared.state.interrupted:
+                break
+            for j, batch in enumerate(dl):
+                # works as a drop_last=True for gradient accumulation
+                if j == max_steps_per_epoch:
+                    break
+                scheduler.apply(optimizer, hypernetwork.step)
+                if scheduler.finished:
+                    break
+                if shared.state.interrupted:
+                    break
 
-        if shared.state.interrupted:
-            break
+                with torch.autocast("cuda"):
+                    x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
+                    if tag_drop_out != 0 or shuffle_tags:
+                        shared.sd_model.cond_stage_model.to(devices.device)
+                        c = shared.sd_model.cond_stage_model(batch.cond_text).to(devices.device, non_blocking=pin_memory)
+                        shared.sd_model.cond_stage_model.to(devices.cpu)
+                    else:
+                        c = stack_conds(batch.cond).to(devices.device, non_blocking=pin_memory)
+                    loss = shared.sd_model(x, c)[0] / gradient_step
+                    del x
+                    del c
 
-        with torch.autocast("cuda"):
-            c = stack_conds([entry.cond for entry in entries]).to(devices.device)
-            # c = torch.vstack([entry.cond for entry in entries]).to(devices.device)
-            x = torch.stack([entry.latent for entry in entries]).to(devices.device)
-            loss = shared.sd_model(x, c)[0]
-            del x
-            del c
+                    _loss_step += loss.item()
+                scaler.scale(loss).backward()
+                # go back until we reach gradient accumulation steps
+                if (j + 1) % gradient_step != 0:
+                    continue
+                # print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.7f}")
+                # scaler.unscale_(optimizer)
+                # print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.15f}")
+                # torch.nn.utils.clip_grad_norm_(weights, max_norm=1.0)
+                # print(f"grad:{weights[0].grad.detach().cpu().abs().mean().item():.15f}")
+                scaler.step(optimizer)
+                scaler.update()
+                hypernetwork.step += 1
+                pbar.update()
+                optimizer.zero_grad(set_to_none=True)
+                loss_step = _loss_step
+                _loss_step = 0
 
-            losses[hypernetwork.step % losses.shape[0]] = loss.item()
-            for entry in entries:
-                loss_dict[entry.filename].append(loss.item())
+                steps_done = hypernetwork.step + 1
                 
-            optimizer.zero_grad()
-            weights[0].grad = None
-            loss.backward()
+                epoch_num = hypernetwork.step // steps_per_epoch
+                epoch_step = hypernetwork.step % steps_per_epoch
 
-            if weights[0].grad is None:
-                steps_without_grad += 1
-            else:
-                steps_without_grad = 0
-            assert steps_without_grad < 10, 'no gradient found for the trained weight after backward() for 10 steps in a row; this is a bug; training cannot continue'
+                pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}")
+                if hypernetwork_dir is not None and steps_done % save_hypernetwork_every == 0:
+                    # Before saving, change name to match current checkpoint.
+                    hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}'
+                    last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name_every}.pt')
+                    hypernetwork.optimizer_name = optimizer_name
+                    if shared.opts.save_optimizer_state:
+                        hypernetwork.optimizer_state_dict = optimizer.state_dict()
+                    save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file)
+                    hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
 
-            optimizer.step()
+                textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, steps_per_epoch, {
+                    "loss": f"{loss_step:.7f}",
+                    "learn_rate": scheduler.learn_rate
+                })
 
-        steps_done = hypernetwork.step + 1
+                if images_dir is not None and steps_done % create_image_every == 0:
+                    forced_filename = f'{hypernetwork_name}-{steps_done}'
+                    last_saved_image = os.path.join(images_dir, forced_filename)
+                    hypernetwork.eval_mode()
+                    shared.sd_model.cond_stage_model.to(devices.device)
+                    shared.sd_model.first_stage_model.to(devices.device)
 
-        if torch.isnan(losses[hypernetwork.step % losses.shape[0]]): 
-            raise RuntimeError("Loss diverged.")
-        
-        if len(previous_mean_losses) > 1:
-            std = stdev(previous_mean_losses)
-        else:
-            std = 0
-        dataset_loss_info = f"dataset loss:{mean(previous_mean_losses):.3f}" + u"\u00B1" + f"({std / (len(previous_mean_losses) ** 0.5):.3f})"
-        pbar.set_description(dataset_loss_info)
+                    p = processing.StableDiffusionProcessingTxt2Img(
+                        sd_model=shared.sd_model,
+                        do_not_save_grid=True,
+                        do_not_save_samples=True,
+                    )
 
-        if hypernetwork_dir is not None and steps_done % save_hypernetwork_every == 0:
-            # Before saving, change name to match current checkpoint.
-            hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}'
-            last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name_every}.pt')
-            hypernetwork.optimizer_name = optimizer_name
-            if shared.opts.save_optimizer_state:
-                hypernetwork.optimizer_state_dict = optimizer.state_dict()
-            save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file)
-            hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
+                    if preview_from_txt2img:
+                        p.prompt = preview_prompt
+                        p.negative_prompt = preview_negative_prompt
+                        p.steps = preview_steps
+                        p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
+                        p.cfg_scale = preview_cfg_scale
+                        p.seed = preview_seed
+                        p.width = preview_width
+                        p.height = preview_height
+                    else:
+                        p.prompt = batch.cond_text[0]
+                        p.steps = 20
+                        p.width = training_width
+                        p.height = training_height
 
-        textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, len(ds), {
-            "loss": f"{previous_mean_loss:.7f}",
-            "learn_rate": scheduler.learn_rate
-        })
+                    preview_text = p.prompt
 
-        if images_dir is not None and steps_done % create_image_every == 0:
-            forced_filename = f'{hypernetwork_name}-{steps_done}'
-            last_saved_image = os.path.join(images_dir, forced_filename)
+                    processed = processing.process_images(p)
+                    image = processed.images[0] if len(processed.images) > 0 else None
 
-            optimizer.zero_grad()
-            shared.sd_model.cond_stage_model.to(devices.device)
-            shared.sd_model.first_stage_model.to(devices.device)
+                    if unload:
+                        shared.sd_model.cond_stage_model.to(devices.cpu)
+                        shared.sd_model.first_stage_model.to(devices.cpu)
+                    hypernetwork.train_mode()
+                    if image is not None:
+                        shared.state.current_image = image
+                        last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
+                        last_saved_image += f", prompt: {preview_text}"
 
-            p = processing.StableDiffusionProcessingTxt2Img(
-                sd_model=shared.sd_model,
-                do_not_save_grid=True,
-                do_not_save_samples=True,
-            )
+                shared.state.job_no = hypernetwork.step
 
-            if preview_from_txt2img:
-                p.prompt = preview_prompt
-                p.negative_prompt = preview_negative_prompt
-                p.steps = preview_steps
-                p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
-                p.cfg_scale = preview_cfg_scale
-                p.seed = preview_seed
-                p.width = preview_width
-                p.height = preview_height
-            else:
-                p.prompt = entries[0].cond_text
-                p.steps = 20
-
-            preview_text = p.prompt
-
-            processed = processing.process_images(p)
-            image = processed.images[0] if len(processed.images)>0 else None
-
-            if unload:
-                shared.sd_model.cond_stage_model.to(devices.cpu)
-                shared.sd_model.first_stage_model.to(devices.cpu)
-
-            if image is not None:
-                shared.state.current_image = image
-                last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
-                last_saved_image += f", prompt: {preview_text}"
-
-        shared.state.job_no = hypernetwork.step
-
-        shared.state.textinfo = f"""
+                shared.state.textinfo = f"""
 <p>
-Loss: {previous_mean_loss:.7f}<br/>
-Step: {hypernetwork.step}<br/>
-Last prompt: {html.escape(entries[0].cond_text)}<br/>
+Loss: {loss_step:.7f}<br/>
+Step: {steps_done}<br/>
+Last prompt: {html.escape(batch.cond_text[0])}<br/>
 Last saved hypernetwork: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
 </p>
 """
-        
-    report_statistics(loss_dict)
+    except Exception:
+        print(traceback.format_exc(), file=sys.stderr)
+    finally:
+        pbar.leave = False
+        pbar.close()
+        hypernetwork.eval_mode()
+        #report_statistics(loss_dict)
 
     filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
     hypernetwork.optimizer_name = optimizer_name
@@ -579,6 +614,9 @@ Last saved image: {html.escape(last_saved_image)}<br/>
     save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename)
     del optimizer
     hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
+    shared.sd_model.cond_stage_model.to(devices.device)
+    shared.sd_model.first_stage_model.to(devices.device)
+
     return hypernetwork, filename
 
 def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename):

modules/images.py

@@ -524,6 +524,8 @@ def save_image(image, path, basename, seed=None, prompt=None, extension='png', i
     else:
         image.save(fullfn, quality=opts.jpeg_quality)
 
+    image.already_saved_as = fullfn
+
     target_side_length = 4000
     oversize = image.width > target_side_length or image.height > target_side_length
     if opts.export_for_4chan and (oversize or os.stat(fullfn).st_size > 4 * 1024 * 1024):

modules/lowvram.py

@@ -51,6 +51,10 @@ def setup_for_low_vram(sd_model, use_medvram):
         send_me_to_gpu(first_stage_model, None)
         return first_stage_model_decode(z)
 
+    # for SD1, cond_stage_model is CLIP and its NN is in the tranformer frield, but for SD2, it's open clip, and it's in model field
+    if hasattr(sd_model.cond_stage_model, 'model'):
+        sd_model.cond_stage_model.transformer = sd_model.cond_stage_model.model
+
     # remove three big modules, cond, first_stage, and unet from the model and then
     # send the model to GPU. Then put modules back. the modules will be in CPU.
     stored = sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.model
@@ -65,6 +69,10 @@ def setup_for_low_vram(sd_model, use_medvram):
     sd_model.first_stage_model.decode = first_stage_model_decode_wrap
     parents[sd_model.cond_stage_model.transformer] = sd_model.cond_stage_model
 
+    if hasattr(sd_model.cond_stage_model, 'model'):
+        sd_model.cond_stage_model.model = sd_model.cond_stage_model.transformer
+        del sd_model.cond_stage_model.transformer
+
     if use_medvram:
         sd_model.model.register_forward_pre_hook(send_me_to_gpu)
     else:

modules/paths.py

@@ -9,7 +9,7 @@ sys.path.insert(0, script_path)
 
 # search for directory of stable diffusion in following places
 sd_path = None
-possible_sd_paths = [os.path.join(script_path, 'repositories/stable-diffusion'), '.', os.path.dirname(script_path)]
+possible_sd_paths = [os.path.join(script_path, 'repositories/stable-diffusion-stability-ai'), '.', os.path.dirname(script_path)]
 for possible_sd_path in possible_sd_paths:
     if os.path.exists(os.path.join(possible_sd_path, 'ldm/models/diffusion/ddpm.py')):
         sd_path = os.path.abspath(possible_sd_path)

modules/sd_hijack.py

@@ -8,19 +8,31 @@ from torch import einsum
 from torch.nn.functional import silu
 
 import modules.textual_inversion.textual_inversion
-from modules import prompt_parser, devices, sd_hijack_optimizations, shared
+from modules import prompt_parser, devices, sd_hijack_optimizations, shared, sd_hijack_checkpoint
+from modules.hypernetworks import hypernetwork
 from modules.shared import opts, device, cmd_opts
+from modules import sd_hijack_clip, sd_hijack_open_clip
+
 from modules.sd_hijack_optimizations import invokeAI_mps_available
 
 import ldm.modules.attention
 import ldm.modules.diffusionmodules.model
 import ldm.models.diffusion.ddim
 import ldm.models.diffusion.plms
+import ldm.modules.encoders.modules
 
 attention_CrossAttention_forward = ldm.modules.attention.CrossAttention.forward
 diffusionmodules_model_nonlinearity = ldm.modules.diffusionmodules.model.nonlinearity
 diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.AttnBlock.forward
 
+# new memory efficient cross attention blocks do not support hypernets and we already
+# have memory efficient cross attention anyway, so this disables SD2.0's memory efficient cross attention
+ldm.modules.attention.MemoryEfficientCrossAttention = ldm.modules.attention.CrossAttention
+ldm.modules.attention.BasicTransformerBlock.ATTENTION_MODES["softmax-xformers"] = ldm.modules.attention.CrossAttention
+
+# silence new console spam from SD2
+ldm.modules.attention.print = lambda *args: None
+ldm.modules.diffusionmodules.model.print = lambda *args: None
 
 def apply_optimizations():
     undo_optimizations()
@@ -49,16 +61,15 @@ def apply_optimizations():
 
 
 def undo_optimizations():
-    from modules.hypernetworks import hypernetwork
-
     ldm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
     ldm.modules.diffusionmodules.model.nonlinearity = diffusionmodules_model_nonlinearity
     ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
 
 
-def get_target_prompt_token_count(token_count):
-    return math.ceil(max(token_count, 1) / 75) * 75
-
+def fix_checkpoint():
+    ldm.modules.attention.BasicTransformerBlock.forward = sd_hijack_checkpoint.BasicTransformerBlock_forward
+    ldm.modules.diffusionmodules.openaimodel.ResBlock.forward = sd_hijack_checkpoint.ResBlock_forward
+    ldm.modules.diffusionmodules.openaimodel.AttentionBlock.forward = sd_hijack_checkpoint.AttentionBlock_forward
 
 class StableDiffusionModelHijack:
     fixes = None
@@ -70,14 +81,18 @@ class StableDiffusionModelHijack:
     embedding_db = modules.textual_inversion.textual_inversion.EmbeddingDatabase(cmd_opts.embeddings_dir)
 
     def hijack(self, m):
-        model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
-
-        model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
-        m.cond_stage_model = FrozenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
+        if type(m.cond_stage_model) == ldm.modules.encoders.modules.FrozenCLIPEmbedder:
+            model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
+            model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
+            m.cond_stage_model = sd_hijack_clip.FrozenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
+        elif type(m.cond_stage_model) == ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder:
+            m.cond_stage_model.model.token_embedding = EmbeddingsWithFixes(m.cond_stage_model.model.token_embedding, self)
+            m.cond_stage_model = sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords(m.cond_stage_model, self)
 
         self.clip = m.cond_stage_model
 
         apply_optimizations()
+        fix_checkpoint()
 
         def flatten(el):
             flattened = [flatten(children) for children in el.children()]
@@ -89,12 +104,15 @@ class StableDiffusionModelHijack:
         self.layers = flatten(m)
 
     def undo_hijack(self, m):
-        if type(m.cond_stage_model) == FrozenCLIPEmbedderWithCustomWords:
+        if type(m.cond_stage_model) == sd_hijack_clip.FrozenCLIPEmbedderWithCustomWords:
             m.cond_stage_model = m.cond_stage_model.wrapped
 
-        model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
-        if type(model_embeddings.token_embedding) == EmbeddingsWithFixes:
-            model_embeddings.token_embedding = model_embeddings.token_embedding.wrapped
+            model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
+            if type(model_embeddings.token_embedding) == EmbeddingsWithFixes:
+                model_embeddings.token_embedding = model_embeddings.token_embedding.wrapped
+        elif type(m.cond_stage_model) == sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords:
+            m.cond_stage_model.wrapped.model.token_embedding = m.cond_stage_model.wrapped.model.token_embedding.wrapped
+            m.cond_stage_model = m.cond_stage_model.wrapped
 
         self.apply_circular(False)
         self.layers = None
@@ -114,262 +132,9 @@ class StableDiffusionModelHijack:
 
     def tokenize(self, text):
         _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
-        return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count)
+        return remade_batch_tokens[0], token_count, sd_hijack_clip.get_target_prompt_token_count(token_count)
 
 
-class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
-    def __init__(self, wrapped, hijack):
-        super().__init__()
-        self.wrapped = wrapped
-        self.hijack: StableDiffusionModelHijack = hijack
-        self.tokenizer = wrapped.tokenizer
-        self.token_mults = {}
-
-        self.comma_token = [v for k, v in self.tokenizer.get_vocab().items() if k == ',</w>'][0]
-
-        tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if '(' in k or ')' in k or '[' in k or ']' in k]
-        for text, ident in tokens_with_parens:
-            mult = 1.0
-            for c in text:
-                if c == '[':
-                    mult /= 1.1
-                if c == ']':
-                    mult *= 1.1
-                if c == '(':
-                    mult *= 1.1
-                if c == ')':
-                    mult /= 1.1
-
-            if mult != 1.0:
-                self.token_mults[ident] = mult
-
-    def tokenize_line(self, line, used_custom_terms, hijack_comments):
-        id_end = self.wrapped.tokenizer.eos_token_id
-
-        if opts.enable_emphasis:
-            parsed = prompt_parser.parse_prompt_attention(line)
-        else:
-            parsed = [[line, 1.0]]
-
-        tokenized = self.wrapped.tokenizer([text for text, _ in parsed], truncation=False, add_special_tokens=False)["input_ids"]
-
-        fixes = []
-        remade_tokens = []
-        multipliers = []
-        last_comma = -1
-
-        for tokens, (text, weight) in zip(tokenized, parsed):
-            i = 0
-            while i < len(tokens):
-                token = tokens[i]
-
-                embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
-
-                if token == self.comma_token:
-                    last_comma = len(remade_tokens)
-                elif opts.comma_padding_backtrack != 0 and max(len(remade_tokens), 1) % 75 == 0 and last_comma != -1 and len(remade_tokens) - last_comma <= opts.comma_padding_backtrack:
-                    last_comma += 1
-                    reloc_tokens = remade_tokens[last_comma:]
-                    reloc_mults = multipliers[last_comma:]
-
-                    remade_tokens = remade_tokens[:last_comma]
-                    length = len(remade_tokens)
-
-                    rem = int(math.ceil(length / 75)) * 75 - length
-                    remade_tokens += [id_end] * rem + reloc_tokens
-                    multipliers = multipliers[:last_comma] + [1.0] * rem + reloc_mults
-
-                if embedding is None:
-                    remade_tokens.append(token)
-                    multipliers.append(weight)
-                    i += 1
-                else:
-                    emb_len = int(embedding.vec.shape[0])
-                    iteration = len(remade_tokens) // 75
-                    if (len(remade_tokens) + emb_len) // 75 != iteration:
-                        rem = (75 * (iteration + 1) - len(remade_tokens))
-                        remade_tokens += [id_end] * rem
-                        multipliers += [1.0] * rem
-                        iteration += 1
-                    fixes.append((iteration, (len(remade_tokens) % 75, embedding)))
-                    remade_tokens += [0] * emb_len
-                    multipliers += [weight] * emb_len
-                    used_custom_terms.append((embedding.name, embedding.checksum()))
-                    i += embedding_length_in_tokens
-
-        token_count = len(remade_tokens)
-        prompt_target_length = get_target_prompt_token_count(token_count)
-        tokens_to_add = prompt_target_length - len(remade_tokens)
-
-        remade_tokens = remade_tokens + [id_end] * tokens_to_add
-        multipliers = multipliers + [1.0] * tokens_to_add
-
-        return remade_tokens, fixes, multipliers, token_count
-
-    def process_text(self, texts):
-        used_custom_terms = []
-        remade_batch_tokens = []
-        hijack_comments = []
-        hijack_fixes = []
-        token_count = 0
-
-        cache = {}
-        batch_multipliers = []
-        for line in texts:
-            if line in cache:
-                remade_tokens, fixes, multipliers = cache[line]
-            else:
-                remade_tokens, fixes, multipliers, current_token_count = self.tokenize_line(line, used_custom_terms, hijack_comments)
-                token_count = max(current_token_count, token_count)
-
-                cache[line] = (remade_tokens, fixes, multipliers)
-
-            remade_batch_tokens.append(remade_tokens)
-            hijack_fixes.append(fixes)
-            batch_multipliers.append(multipliers)
-
-        return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
-
-    def process_text_old(self, text):
-        id_start = self.wrapped.tokenizer.bos_token_id
-        id_end = self.wrapped.tokenizer.eos_token_id
-        maxlen = self.wrapped.max_length  # you get to stay at 77
-        used_custom_terms = []
-        remade_batch_tokens = []
-        overflowing_words = []
-        hijack_comments = []
-        hijack_fixes = []
-        token_count = 0
-
-        cache = {}
-        batch_tokens = self.wrapped.tokenizer(text, truncation=False, add_special_tokens=False)["input_ids"]
-        batch_multipliers = []
-        for tokens in batch_tokens:
-            tuple_tokens = tuple(tokens)
-
-            if tuple_tokens in cache:
-                remade_tokens, fixes, multipliers = cache[tuple_tokens]
-            else:
-                fixes = []
-                remade_tokens = []
-                multipliers = []
-                mult = 1.0
-
-                i = 0
-                while i < len(tokens):
-                    token = tokens[i]
-
-                    embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
-
-                    mult_change = self.token_mults.get(token) if opts.enable_emphasis else None
-                    if mult_change is not None:
-                        mult *= mult_change
-                        i += 1
-                    elif embedding is None:
-                        remade_tokens.append(token)
-                        multipliers.append(mult)
-                        i += 1
-                    else:
-                        emb_len = int(embedding.vec.shape[0])
-                        fixes.append((len(remade_tokens), embedding))
-                        remade_tokens += [0] * emb_len
-                        multipliers += [mult] * emb_len
-                        used_custom_terms.append((embedding.name, embedding.checksum()))
-                        i += embedding_length_in_tokens
-
-                if len(remade_tokens) > maxlen - 2:
-                    vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
-                    ovf = remade_tokens[maxlen - 2:]
-                    overflowing_words = [vocab.get(int(x), "") for x in ovf]
-                    overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
-                    hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
-
-                token_count = len(remade_tokens)
-                remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
-                remade_tokens = [id_start] + remade_tokens[0:maxlen - 2] + [id_end]
-                cache[tuple_tokens] = (remade_tokens, fixes, multipliers)
-
-            multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
-            multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
-
-            remade_batch_tokens.append(remade_tokens)
-            hijack_fixes.append(fixes)
-            batch_multipliers.append(multipliers)
-        return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
-
-    def forward(self, text):
-        use_old = opts.use_old_emphasis_implementation
-        if use_old:
-            batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text_old(text)
-        else:
-            batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text(text)
-
-        self.hijack.comments += hijack_comments
-
-        if len(used_custom_terms) > 0:
-            self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
-
-        if use_old:
-            self.hijack.fixes = hijack_fixes
-            return self.process_tokens(remade_batch_tokens, batch_multipliers)
-
-        z = None
-        i = 0
-        while max(map(len, remade_batch_tokens)) != 0:
-            rem_tokens = [x[75:] for x in remade_batch_tokens]
-            rem_multipliers = [x[75:] for x in batch_multipliers]
-
-            self.hijack.fixes = []
-            for unfiltered in hijack_fixes:
-                fixes = []
-                for fix in unfiltered:
-                    if fix[0] == i:
-                        fixes.append(fix[1])
-                self.hijack.fixes.append(fixes)
-
-            tokens = []
-            multipliers = []
-            for j in range(len(remade_batch_tokens)):
-                if len(remade_batch_tokens[j]) > 0:
-                    tokens.append(remade_batch_tokens[j][:75])
-                    multipliers.append(batch_multipliers[j][:75])
-                else:
-                    tokens.append([self.wrapped.tokenizer.eos_token_id] * 75)
-                    multipliers.append([1.0] * 75)
-
-            z1 = self.process_tokens(tokens, multipliers)
-            z = z1 if z is None else torch.cat((z, z1), axis=-2)
-
-            remade_batch_tokens = rem_tokens
-            batch_multipliers = rem_multipliers
-            i += 1
-
-        return z
-
-    def process_tokens(self, remade_batch_tokens, batch_multipliers):
-        if not opts.use_old_emphasis_implementation:
-            remade_batch_tokens = [[self.wrapped.tokenizer.bos_token_id] + x[:75] + [self.wrapped.tokenizer.eos_token_id] for x in remade_batch_tokens]
-            batch_multipliers = [[1.0] + x[:75] + [1.0] for x in batch_multipliers]
-
-        tokens = torch.asarray(remade_batch_tokens).to(device)
-        outputs = self.wrapped.transformer(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers)
-
-        if opts.CLIP_stop_at_last_layers > 1:
-            z = outputs.hidden_states[-opts.CLIP_stop_at_last_layers]
-            z = self.wrapped.transformer.text_model.final_layer_norm(z)
-        else:
-            z = outputs.last_hidden_state
-
-        # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
-        batch_multipliers_of_same_length = [x + [1.0] * (75 - len(x)) for x in batch_multipliers]
-        batch_multipliers = torch.asarray(batch_multipliers_of_same_length).to(device)
-        original_mean = z.mean()
-        z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
-        new_mean = z.mean()
-        z *= original_mean / new_mean
-
-        return z
-
 
 class EmbeddingsWithFixes(torch.nn.Module):
     def __init__(self, wrapped, embeddings):

modules/sd_hijack_checkpoint.py

@@ -0,0 +1,10 @@
+from torch.utils.checkpoint import checkpoint
+
+def BasicTransformerBlock_forward(self, x, context=None):
+    return checkpoint(self._forward, x, context)
+
+def AttentionBlock_forward(self, x):
+    return checkpoint(self._forward, x)
+
+def ResBlock_forward(self, x, emb):
+    return checkpoint(self._forward, x, emb)

modules/sd_hijack_clip.py

@@ -0,0 +1,301 @@
+import math
+
+import torch
+
+from modules import prompt_parser, devices
+from modules.shared import opts
+
+
+def get_target_prompt_token_count(token_count):
+    return math.ceil(max(token_count, 1) / 75) * 75
+
+
+class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module):
+    def __init__(self, wrapped, hijack):
+        super().__init__()
+        self.wrapped = wrapped
+        self.hijack = hijack
+
+    def tokenize(self, texts):
+        raise NotImplementedError
+
+    def encode_with_transformers(self, tokens):
+        raise NotImplementedError
+
+    def encode_embedding_init_text(self, init_text, nvpt):
+        raise NotImplementedError
+
+    def tokenize_line(self, line, used_custom_terms, hijack_comments):
+        if opts.enable_emphasis:
+            parsed = prompt_parser.parse_prompt_attention(line)
+        else:
+            parsed = [[line, 1.0]]
+
+        tokenized = self.tokenize([text for text, _ in parsed])
+
+        fixes = []
+        remade_tokens = []
+        multipliers = []
+        last_comma = -1
+
+        for tokens, (text, weight) in zip(tokenized, parsed):
+            i = 0
+            while i < len(tokens):
+                token = tokens[i]
+
+                embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
+
+                if token == self.comma_token:
+                    last_comma = len(remade_tokens)
+                elif opts.comma_padding_backtrack != 0 and max(len(remade_tokens), 1) % 75 == 0 and last_comma != -1 and len(remade_tokens) - last_comma <= opts.comma_padding_backtrack:
+                    last_comma += 1
+                    reloc_tokens = remade_tokens[last_comma:]
+                    reloc_mults = multipliers[last_comma:]
+
+                    remade_tokens = remade_tokens[:last_comma]
+                    length = len(remade_tokens)
+
+                    rem = int(math.ceil(length / 75)) * 75 - length
+                    remade_tokens += [self.id_end] * rem + reloc_tokens
+                    multipliers = multipliers[:last_comma] + [1.0] * rem + reloc_mults
+
+                if embedding is None:
+                    remade_tokens.append(token)
+                    multipliers.append(weight)
+                    i += 1
+                else:
+                    emb_len = int(embedding.vec.shape[0])
+                    iteration = len(remade_tokens) // 75
+                    if (len(remade_tokens) + emb_len) // 75 != iteration:
+                        rem = (75 * (iteration + 1) - len(remade_tokens))
+                        remade_tokens += [self.id_end] * rem
+                        multipliers += [1.0] * rem
+                        iteration += 1
+                    fixes.append((iteration, (len(remade_tokens) % 75, embedding)))
+                    remade_tokens += [0] * emb_len
+                    multipliers += [weight] * emb_len
+                    used_custom_terms.append((embedding.name, embedding.checksum()))
+                    i += embedding_length_in_tokens
+
+        token_count = len(remade_tokens)
+        prompt_target_length = get_target_prompt_token_count(token_count)
+        tokens_to_add = prompt_target_length - len(remade_tokens)
+
+        remade_tokens = remade_tokens + [self.id_end] * tokens_to_add
+        multipliers = multipliers + [1.0] * tokens_to_add
+
+        return remade_tokens, fixes, multipliers, token_count
+
+    def process_text(self, texts):
+        used_custom_terms = []
+        remade_batch_tokens = []
+        hijack_comments = []
+        hijack_fixes = []
+        token_count = 0
+
+        cache = {}
+        batch_multipliers = []
+        for line in texts:
+            if line in cache:
+                remade_tokens, fixes, multipliers = cache[line]
+            else:
+                remade_tokens, fixes, multipliers, current_token_count = self.tokenize_line(line, used_custom_terms, hijack_comments)
+                token_count = max(current_token_count, token_count)
+
+                cache[line] = (remade_tokens, fixes, multipliers)
+
+            remade_batch_tokens.append(remade_tokens)
+            hijack_fixes.append(fixes)
+            batch_multipliers.append(multipliers)
+
+        return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
+
+    def process_text_old(self, texts):
+        id_start = self.id_start
+        id_end = self.id_end
+        maxlen = self.wrapped.max_length  # you get to stay at 77
+        used_custom_terms = []
+        remade_batch_tokens = []
+        hijack_comments = []
+        hijack_fixes = []
+        token_count = 0
+
+        cache = {}
+        batch_tokens = self.tokenize(texts)
+        batch_multipliers = []
+        for tokens in batch_tokens:
+            tuple_tokens = tuple(tokens)
+
+            if tuple_tokens in cache:
+                remade_tokens, fixes, multipliers = cache[tuple_tokens]
+            else:
+                fixes = []
+                remade_tokens = []
+                multipliers = []
+                mult = 1.0
+
+                i = 0
+                while i < len(tokens):
+                    token = tokens[i]
+
+                    embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
+
+                    mult_change = self.token_mults.get(token) if opts.enable_emphasis else None
+                    if mult_change is not None:
+                        mult *= mult_change
+                        i += 1
+                    elif embedding is None:
+                        remade_tokens.append(token)
+                        multipliers.append(mult)
+                        i += 1
+                    else:
+                        emb_len = int(embedding.vec.shape[0])
+                        fixes.append((len(remade_tokens), embedding))
+                        remade_tokens += [0] * emb_len
+                        multipliers += [mult] * emb_len
+                        used_custom_terms.append((embedding.name, embedding.checksum()))
+                        i += embedding_length_in_tokens
+
+                if len(remade_tokens) > maxlen - 2:
+                    vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
+                    ovf = remade_tokens[maxlen - 2:]
+                    overflowing_words = [vocab.get(int(x), "") for x in ovf]
+                    overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
+                    hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
+
+                token_count = len(remade_tokens)
+                remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
+                remade_tokens = [id_start] + remade_tokens[0:maxlen - 2] + [id_end]
+                cache[tuple_tokens] = (remade_tokens, fixes, multipliers)
+
+            multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
+            multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
+
+            remade_batch_tokens.append(remade_tokens)
+            hijack_fixes.append(fixes)
+            batch_multipliers.append(multipliers)
+        return batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count
+
+    def forward(self, text):
+        use_old = opts.use_old_emphasis_implementation
+        if use_old:
+            batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text_old(text)
+        else:
+            batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text(text)
+
+        self.hijack.comments += hijack_comments
+
+        if len(used_custom_terms) > 0:
+            self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
+
+        if use_old:
+            self.hijack.fixes = hijack_fixes
+            return self.process_tokens(remade_batch_tokens, batch_multipliers)
+
+        z = None
+        i = 0
+        while max(map(len, remade_batch_tokens)) != 0:
+            rem_tokens = [x[75:] for x in remade_batch_tokens]
+            rem_multipliers = [x[75:] for x in batch_multipliers]
+
+            self.hijack.fixes = []
+            for unfiltered in hijack_fixes:
+                fixes = []
+                for fix in unfiltered:
+                    if fix[0] == i:
+                        fixes.append(fix[1])
+                self.hijack.fixes.append(fixes)
+
+            tokens = []
+            multipliers = []
+            for j in range(len(remade_batch_tokens)):
+                if len(remade_batch_tokens[j]) > 0:
+                    tokens.append(remade_batch_tokens[j][:75])
+                    multipliers.append(batch_multipliers[j][:75])
+                else:
+                    tokens.append([self.id_end] * 75)
+                    multipliers.append([1.0] * 75)
+
+            z1 = self.process_tokens(tokens, multipliers)
+            z = z1 if z is None else torch.cat((z, z1), axis=-2)
+
+            remade_batch_tokens = rem_tokens
+            batch_multipliers = rem_multipliers
+            i += 1
+
+        return z
+
+    def process_tokens(self, remade_batch_tokens, batch_multipliers):
+        if not opts.use_old_emphasis_implementation:
+            remade_batch_tokens = [[self.id_start] + x[:75] + [self.id_end] for x in remade_batch_tokens]
+            batch_multipliers = [[1.0] + x[:75] + [1.0] for x in batch_multipliers]
+
+        tokens = torch.asarray(remade_batch_tokens).to(devices.device)
+
+        if self.id_end != self.id_pad:
+            for batch_pos in range(len(remade_batch_tokens)):
+                index = remade_batch_tokens[batch_pos].index(self.id_end)
+                tokens[batch_pos, index+1:tokens.shape[1]] = self.id_pad
+
+        z = self.encode_with_transformers(tokens)
+
+        # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
+        batch_multipliers_of_same_length = [x + [1.0] * (75 - len(x)) for x in batch_multipliers]
+        batch_multipliers = torch.asarray(batch_multipliers_of_same_length).to(devices.device)
+        original_mean = z.mean()
+        z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
+        new_mean = z.mean()
+        z *= original_mean / new_mean
+
+        return z
+
+
+class FrozenCLIPEmbedderWithCustomWords(FrozenCLIPEmbedderWithCustomWordsBase):
+    def __init__(self, wrapped, hijack):
+        super().__init__(wrapped, hijack)
+        self.tokenizer = wrapped.tokenizer
+        self.comma_token = [v for k, v in self.tokenizer.get_vocab().items() if k == ',</w>'][0]
+
+        self.token_mults = {}
+        tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if '(' in k or ')' in k or '[' in k or ']' in k]
+        for text, ident in tokens_with_parens:
+            mult = 1.0
+            for c in text:
+                if c == '[':
+                    mult /= 1.1
+                if c == ']':
+                    mult *= 1.1
+                if c == '(':
+                    mult *= 1.1
+                if c == ')':
+                    mult /= 1.1
+
+            if mult != 1.0:
+                self.token_mults[ident] = mult
+
+        self.id_start = self.wrapped.tokenizer.bos_token_id
+        self.id_end = self.wrapped.tokenizer.eos_token_id
+        self.id_pad = self.id_end
+
+    def tokenize(self, texts):
+        tokenized = self.wrapped.tokenizer(texts, truncation=False, add_special_tokens=False)["input_ids"]
+
+        return tokenized
+
+    def encode_with_transformers(self, tokens):
+        outputs = self.wrapped.transformer(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers)
+
+        if opts.CLIP_stop_at_last_layers > 1:
+            z = outputs.hidden_states[-opts.CLIP_stop_at_last_layers]
+            z = self.wrapped.transformer.text_model.final_layer_norm(z)
+        else:
+            z = outputs.last_hidden_state
+
+        return z
+
+    def encode_embedding_init_text(self, init_text, nvpt):
+        embedding_layer = self.wrapped.transformer.text_model.embeddings
+        ids = self.wrapped.tokenizer(init_text, max_length=nvpt, return_tensors="pt", add_special_tokens=False)["input_ids"]
+        embedded = embedding_layer.token_embedding.wrapped(ids.to(devices.device)).squeeze(0)
+
+        return embedded

modules/sd_hijack_inpainting.py

@@ -199,8 +199,8 @@ def sample_plms(self,
 
 @torch.no_grad()
 def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
-                    temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
-                    unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None):
+                  temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                  unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None, dynamic_threshold=None):
     b, *_, device = *x.shape, x.device
 
     def get_model_output(x, t):
@@ -249,6 +249,8 @@ def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=F
         pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
         if quantize_denoised:
             pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+        if dynamic_threshold is not None:
+            pred_x0 = norm_thresholding(pred_x0, dynamic_threshold)
         # direction pointing to x_t
         dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
         noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
@@ -321,12 +323,16 @@ def should_hijack_inpainting(checkpoint_info):
 
 
 def do_inpainting_hijack():
-    ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning
+    # most of this stuff seems to no longer be needed because it is already included into SD2.0
+    # LatentInpaintDiffusion remains because SD2.0's LatentInpaintDiffusion can't be loaded without specifying a checkpoint
+    # p_sample_plms is needed because PLMS can't work with dicts as conditionings
+    # this file should be cleaned up later if weverything tuens out to work fine
+
+    # ldm.models.diffusion.ddpm.get_unconditional_conditioning = get_unconditional_conditioning
     ldm.models.diffusion.ddpm.LatentInpaintDiffusion = LatentInpaintDiffusion
 
-    ldm.models.diffusion.ddim.DDIMSampler.p_sample_ddim = p_sample_ddim
-    ldm.models.diffusion.ddim.DDIMSampler.sample = sample_ddim
+    # ldm.models.diffusion.ddim.DDIMSampler.p_sample_ddim = p_sample_ddim
+    # ldm.models.diffusion.ddim.DDIMSampler.sample = sample_ddim
 
     ldm.models.diffusion.plms.PLMSSampler.p_sample_plms = p_sample_plms
-    ldm.models.diffusion.plms.PLMSSampler.sample = sample_plms
-
+    # ldm.models.diffusion.plms.PLMSSampler.sample = sample_plms

modules/sd_hijack_open_clip.py

@@ -0,0 +1,37 @@
+import open_clip.tokenizer
+import torch
+
+from modules import sd_hijack_clip, devices
+from modules.shared import opts
+
+tokenizer = open_clip.tokenizer._tokenizer
+
+
+class FrozenOpenCLIPEmbedderWithCustomWords(sd_hijack_clip.FrozenCLIPEmbedderWithCustomWordsBase):
+    def __init__(self, wrapped, hijack):
+        super().__init__(wrapped, hijack)
+
+        self.comma_token = [v for k, v in tokenizer.encoder.items() if k == ',</w>'][0]
+        self.id_start = tokenizer.encoder["<start_of_text>"]
+        self.id_end = tokenizer.encoder["<end_of_text>"]
+        self.id_pad = 0
+
+    def tokenize(self, texts):
+        assert not opts.use_old_emphasis_implementation, 'Old emphasis implementation not supported for Open Clip'
+
+        tokenized = [tokenizer.encode(text) for text in texts]
+
+        return tokenized
+
+    def encode_with_transformers(self, tokens):
+        # set self.wrapped.layer_idx here according to opts.CLIP_stop_at_last_layers
+        z = self.wrapped.encode_with_transformer(tokens)
+
+        return z
+
+    def encode_embedding_init_text(self, init_text, nvpt):
+        ids = tokenizer.encode(init_text)
+        ids = torch.asarray([ids], device=devices.device, dtype=torch.int)
+        embedded = self.wrapped.model.token_embedding.wrapped(ids).squeeze(0)
+
+        return embedded

modules/sd_samplers.py

@@ -129,7 +129,8 @@ class InterruptedException(BaseException):
 class VanillaStableDiffusionSampler:
     def __init__(self, constructor, sd_model):
         self.sampler = constructor(sd_model)
-        self.orig_p_sample_ddim = self.sampler.p_sample_ddim if hasattr(self.sampler, 'p_sample_ddim') else self.sampler.p_sample_plms
+        self.is_plms = hasattr(self.sampler, 'p_sample_plms')
+        self.orig_p_sample_ddim = self.sampler.p_sample_plms if self.is_plms else self.sampler.p_sample_ddim
         self.mask = None
         self.nmask = None
         self.init_latent = None
@@ -220,7 +221,6 @@ class VanillaStableDiffusionSampler:
         self.mask = p.mask if hasattr(p, 'mask') else None
         self.nmask = p.nmask if hasattr(p, 'nmask') else None
 
-
     def adjust_steps_if_invalid(self, p, num_steps):
         if  (self.config.name == 'DDIM' and p.ddim_discretize == 'uniform') or (self.config.name == 'PLMS'):
             valid_step = 999 / (1000 // num_steps)
@@ -229,7 +229,6 @@ class VanillaStableDiffusionSampler:
         
         return num_steps
 
-
     def sample_img2img(self, p, x, noise, conditioning, unconditional_conditioning, steps=None, image_conditioning=None):
         steps, t_enc = setup_img2img_steps(p, steps)
         steps = self.adjust_steps_if_invalid(p, steps)
@@ -262,9 +261,10 @@ class VanillaStableDiffusionSampler:
         steps = self.adjust_steps_if_invalid(p, steps or p.steps)
 
         # Wrap the conditioning models with additional image conditioning for inpainting model
+        # dummy_for_plms is needed because PLMS code checks the first item in the dict to have the right shape
         if image_conditioning is not None:
-            conditioning = {"c_concat": [image_conditioning], "c_crossattn": [conditioning]}
-            unconditional_conditioning = {"c_concat": [image_conditioning], "c_crossattn": [unconditional_conditioning]}
+            conditioning = {"dummy_for_plms": np.zeros((conditioning.shape[0],)), "c_crossattn": [conditioning], "c_concat": [image_conditioning]}
+            unconditional_conditioning = {"c_crossattn": [unconditional_conditioning], "c_concat": [image_conditioning]}
 
         samples_ddim = self.launch_sampling(steps, lambda: self.sampler.sample(S=steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=self.eta)[0])
 
@@ -352,7 +352,9 @@ class TorchHijack:
 
 class KDiffusionSampler:
     def __init__(self, funcname, sd_model):
-        self.model_wrap = k_diffusion.external.CompVisDenoiser(sd_model, quantize=shared.opts.enable_quantization)
+        denoiser = k_diffusion.external.CompVisVDenoiser if sd_model.parameterization == "v" else k_diffusion.external.CompVisDenoiser
+
+        self.model_wrap = denoiser(sd_model, quantize=shared.opts.enable_quantization)
         self.funcname = funcname
         self.func = getattr(k_diffusion.sampling, self.funcname)
         self.extra_params = sampler_extra_params.get(funcname, [])

modules/shared.py

@@ -11,17 +11,18 @@ import tqdm
 import modules.artists
 import modules.interrogate
 import modules.memmon
-import modules.sd_models
 import modules.styles
 import modules.devices as devices
-from modules import sd_samplers, sd_models, localization, sd_vae, extensions, script_loading
-from modules.hypernetworks import hypernetwork
+from modules import localization, sd_vae, extensions, script_loading
 from modules.paths import models_path, script_path, sd_path
 
+
+demo = None
+
 sd_model_file = os.path.join(script_path, 'model.ckpt')
 default_sd_model_file = sd_model_file
 parser = argparse.ArgumentParser()
-parser.add_argument("--config", type=str, default=os.path.join(sd_path, "configs/stable-diffusion/v1-inference.yaml"), help="path to config which constructs model",)
+parser.add_argument("--config", type=str, default=os.path.join(script_path, "v1-inference.yaml"), help="path to config which constructs model",)
 parser.add_argument("--ckpt", type=str, default=sd_model_file, help="path to checkpoint of stable diffusion model; if specified, this checkpoint will be added to the list of checkpoints and loaded",)
 parser.add_argument("--ckpt-dir", type=str, default=None, help="Path to directory with stable diffusion checkpoints")
 parser.add_argument("--gfpgan-dir", type=str, help="GFPGAN directory", default=('./src/gfpgan' if os.path.exists('./src/gfpgan') else './GFPGAN'))
@@ -121,10 +122,12 @@ xformers_available = False
 config_filename = cmd_opts.ui_settings_file
 
 os.makedirs(cmd_opts.hypernetwork_dir, exist_ok=True)
-hypernetworks = hypernetwork.list_hypernetworks(cmd_opts.hypernetwork_dir)
+hypernetworks = {}
 loaded_hypernetwork = None
 
+
 def reload_hypernetworks():
+    from modules.hypernetworks import hypernetwork
     global hypernetworks
 
     hypernetworks = hypernetwork.list_hypernetworks(cmd_opts.hypernetwork_dir)
@@ -206,10 +209,11 @@ class State:
         if self.current_latent is None:
             return
 
+        import modules.sd_samplers
         if opts.show_progress_grid:
-            self.current_image = sd_samplers.samples_to_image_grid(self.current_latent)
+            self.current_image = modules.sd_samplers.samples_to_image_grid(self.current_latent)
         else:
-            self.current_image = sd_samplers.sample_to_image(self.current_latent)
+            self.current_image = modules.sd_samplers.sample_to_image(self.current_latent)
 
         self.current_image_sampling_step = self.sampling_step
 
@@ -248,6 +252,21 @@ def options_section(section_identifier, options_dict):
     return options_dict
 
 
+def list_checkpoint_tiles():
+    import modules.sd_models
+    return modules.sd_models.checkpoint_tiles()
+
+
+def refresh_checkpoints():
+    import modules.sd_models
+    return modules.sd_models.list_models()
+
+
+def list_samplers():
+    import modules.sd_samplers
+    return modules.sd_samplers.all_samplers
+
+
 hide_dirs = {"visible": not cmd_opts.hide_ui_dir_config}
 
 options_templates = {}
@@ -276,6 +295,10 @@ options_templates.update(options_section(('saving-images', "Saving images/grids"
     "use_original_name_batch": OptionInfo(False, "Use original name for output filename during batch process in extras tab"),
     "save_selected_only": OptionInfo(True, "When using 'Save' button, only save a single selected image"),
     "do_not_add_watermark": OptionInfo(False, "Do not add watermark to images"),
+
+    "temp_dir":  OptionInfo("", "Directory for temporary images; leave empty for default"),
+    "clean_temp_dir_at_start": OptionInfo(False, "Cleanup non-default temporary directory when starting webui"),
+
 }))
 
 options_templates.update(options_section(('saving-paths', "Paths for saving"), {
@@ -322,8 +345,7 @@ options_templates.update(options_section(('system', "System"), {
 
 options_templates.update(options_section(('training', "Training"), {
     "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."),
-    "shuffle_tags": OptionInfo(False, "Shuffleing tags by ',' when create texts."),
-    "tag_drop_out": OptionInfo(0, "Dropout tags when create texts", gr.Slider, {"minimum": 0, "maximum": 1, "step": 0.1}),
+    "pin_memory": OptionInfo(False, "Turn on pin_memory for DataLoader. Makes training slightly faster but can increase memory usage."),
     "save_optimizer_state": OptionInfo(False, "Saves Optimizer state as separate *.optim file. Training can be resumed with HN itself and matching optim file."),
     "dataset_filename_word_regex": OptionInfo("", "Filename word regex"),
     "dataset_filename_join_string": OptionInfo(" ", "Filename join string"),
@@ -333,7 +355,7 @@ options_templates.update(options_section(('training', "Training"), {
 }))
 
 options_templates.update(options_section(('sd', "Stable Diffusion"), {
-    "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles()}, refresh=sd_models.list_models),
+    "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": list_checkpoint_tiles()}, refresh=refresh_checkpoints),
     "sd_checkpoint_cache": OptionInfo(0, "Checkpoints to cache in RAM", gr.Slider, {"minimum": 0, "maximum": 10, "step": 1}),
     "sd_vae": OptionInfo("auto", "SD VAE", gr.Dropdown, lambda: {"choices": sd_vae.vae_list}, refresh=sd_vae.refresh_vae_list),
     "sd_vae_as_default": OptionInfo(False, "Ignore selected VAE for stable diffusion checkpoints that have their own .vae.pt next to them"),
@@ -385,7 +407,7 @@ options_templates.update(options_section(('ui', "User interface"), {
 }))
 
 options_templates.update(options_section(('sampler-params', "Sampler parameters"), {
-    "hide_samplers": OptionInfo([], "Hide samplers in user interface (requires restart)", gr.CheckboxGroup, lambda: {"choices": [x.name for x in sd_samplers.all_samplers]}),
+    "hide_samplers": OptionInfo([], "Hide samplers in user interface (requires restart)", gr.CheckboxGroup, lambda: {"choices": [x.name for x in list_samplers()]}),
     "eta_ddim": OptionInfo(0.0, "eta (noise multiplier) for DDIM", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
     "eta_ancestral": OptionInfo(1.0, "eta (noise multiplier) for ancestral samplers", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}),
     "ddim_discretize": OptionInfo('uniform', "img2img DDIM discretize", gr.Radio, {"choices": ['uniform', 'quad']}),

modules/textual_inversion/dataset.py

@@ -3,7 +3,7 @@ import numpy as np
 import PIL
 import torch
 from PIL import Image
-from torch.utils.data import Dataset
+from torch.utils.data import Dataset, DataLoader
 from torchvision import transforms
 
 import random
@@ -11,25 +11,28 @@ import tqdm
 from modules import devices, shared
 import re
 
+from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+
 re_numbers_at_start = re.compile(r"^[-\d]+\s*")
 
 
 class DatasetEntry:
-    def __init__(self, filename=None, latent=None, filename_text=None):
+    def __init__(self, filename=None, filename_text=None, latent_dist=None, latent_sample=None, cond=None, cond_text=None, pixel_values=None):
         self.filename = filename
-        self.latent = latent
         self.filename_text = filename_text
-        self.cond = None
-        self.cond_text = None
+        self.latent_dist = latent_dist
+        self.latent_sample = latent_sample
+        self.cond = cond
+        self.cond_text = cond_text
+        self.pixel_values = pixel_values
 
 
 class PersonalizedBase(Dataset):
-    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, device=None, template_file=None, include_cond=False, batch_size=1):
+    def __init__(self, data_root, width, height, repeats, flip_p=0.5, placeholder_token="*", model=None, cond_model=None, device=None, template_file=None, include_cond=False, batch_size=1, gradient_step=1, shuffle_tags=False, tag_drop_out=0, latent_sampling_method='once'):        
         re_word = re.compile(shared.opts.dataset_filename_word_regex) if len(shared.opts.dataset_filename_word_regex) > 0 else None
-
+        
         self.placeholder_token = placeholder_token
 
-        self.batch_size = batch_size
         self.width = width
         self.height = height
         self.flip = transforms.RandomHorizontalFlip(p=flip_p)
@@ -45,11 +48,16 @@ class PersonalizedBase(Dataset):
         assert os.path.isdir(data_root), "Dataset directory doesn't exist"
         assert os.listdir(data_root), "Dataset directory is empty"
 
-        cond_model = shared.sd_model.cond_stage_model
-
         self.image_paths = [os.path.join(data_root, file_path) for file_path in os.listdir(data_root)]
+
+        
+        self.shuffle_tags = shuffle_tags
+        self.tag_drop_out = tag_drop_out
+
         print("Preparing dataset...")
         for path in tqdm.tqdm(self.image_paths):
+            if shared.state.interrupted:
+                raise Exception("inturrupted")
             try:
                 image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC)
             except Exception:
@@ -71,37 +79,49 @@ class PersonalizedBase(Dataset):
             npimage = np.array(image).astype(np.uint8)
             npimage = (npimage / 127.5 - 1.0).astype(np.float32)
 
-            torchdata = torch.from_numpy(npimage).to(device=device, dtype=torch.float32)
-            torchdata = torch.moveaxis(torchdata, 2, 0)
+            torchdata = torch.from_numpy(npimage).permute(2, 0, 1).to(device=device, dtype=torch.float32)
+            latent_sample = None
 
-            init_latent = model.get_first_stage_encoding(model.encode_first_stage(torchdata.unsqueeze(dim=0))).squeeze()
-            init_latent = init_latent.to(devices.cpu)
+            with torch.autocast("cuda"):
+                latent_dist = model.encode_first_stage(torchdata.unsqueeze(dim=0))
 
-            entry = DatasetEntry(filename=path, filename_text=filename_text, latent=init_latent)
+            if latent_sampling_method == "once" or (latent_sampling_method == "deterministic" and not isinstance(latent_dist, DiagonalGaussianDistribution)):
+                latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu)
+                latent_sampling_method = "once"
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample)
+            elif latent_sampling_method == "deterministic":
+                # Works only for DiagonalGaussianDistribution
+                latent_dist.std = 0
+                latent_sample = model.get_first_stage_encoding(latent_dist).squeeze().to(devices.cpu)
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_sample=latent_sample)
+            elif latent_sampling_method == "random":
+                entry = DatasetEntry(filename=path, filename_text=filename_text, latent_dist=latent_dist)
 
-            if include_cond:
+            if not (self.tag_drop_out != 0 or self.shuffle_tags):
                 entry.cond_text = self.create_text(filename_text)
-                entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
+
+            if include_cond and not (self.tag_drop_out != 0 or self.shuffle_tags):
+                with torch.autocast("cuda"):
+                    entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
 
             self.dataset.append(entry)
+            del torchdata
+            del latent_dist
+            del latent_sample
 
-        assert len(self.dataset) > 0, "No images have been found in the dataset."
-        self.length = len(self.dataset) * repeats // batch_size
-
-        self.dataset_length = len(self.dataset)
-        self.indexes = None
-        self.shuffle()
-
-    def shuffle(self):
-        self.indexes = np.random.permutation(self.dataset_length)
+        self.length = len(self.dataset)
+        assert self.length > 0, "No images have been found in the dataset."
+        self.batch_size = min(batch_size, self.length)
+        self.gradient_step = min(gradient_step, self.length // self.batch_size)
+        self.latent_sampling_method = latent_sampling_method
 
     def create_text(self, filename_text):
         text = random.choice(self.lines)
         text = text.replace("[name]", self.placeholder_token)
         tags = filename_text.split(',')
-        if shared.opts.tag_drop_out != 0:
-            tags = [t for t in tags if random.random() > shared.opts.tag_drop_out]
-        if shared.opts.shuffle_tags:
+        if self.tag_drop_out != 0:
+            tags = [t for t in tags if random.random() > self.tag_drop_out]
+        if self.shuffle_tags:
             random.shuffle(tags)
         text = text.replace("[filewords]", ','.join(tags))
         return text
@@ -110,19 +130,43 @@ class PersonalizedBase(Dataset):
         return self.length
 
     def __getitem__(self, i):
-        res = []
+        entry = self.dataset[i]
+        if self.tag_drop_out != 0 or self.shuffle_tags:
+            entry.cond_text = self.create_text(entry.filename_text)
+        if self.latent_sampling_method == "random":
+            entry.latent_sample = shared.sd_model.get_first_stage_encoding(entry.latent_dist).to(devices.cpu)
+        return entry
 
-        for j in range(self.batch_size):
-            position = i * self.batch_size + j
-            if position % len(self.indexes) == 0:
-                self.shuffle()
+class PersonalizedDataLoader(DataLoader):
+    def __init__(self, dataset, latent_sampling_method="once", batch_size=1, pin_memory=False):
+        super(PersonalizedDataLoader, self).__init__(dataset, shuffle=True, drop_last=True, batch_size=batch_size, pin_memory=pin_memory)
+        if latent_sampling_method == "random":
+            self.collate_fn = collate_wrapper_random
+        else:
+            self.collate_fn = collate_wrapper
+        
 
-            index = self.indexes[position % len(self.indexes)]
-            entry = self.dataset[index]
+class BatchLoader:
+    def __init__(self, data):
+        self.cond_text = [entry.cond_text for entry in data]
+        self.cond = [entry.cond for entry in data]
+        self.latent_sample = torch.stack([entry.latent_sample for entry in data]).squeeze(1)
+        #self.emb_index = [entry.emb_index for entry in data]
+        #print(self.latent_sample.device)
 
-            if entry.cond is None:
-                entry.cond_text = self.create_text(entry.filename_text)
+    def pin_memory(self):
+        self.latent_sample = self.latent_sample.pin_memory()
+        return self
 
-            res.append(entry)
+def collate_wrapper(batch):
+    return BatchLoader(batch)
 
-        return res
+class BatchLoaderRandom(BatchLoader):
+    def __init__(self, data):
+        super().__init__(data)
+
+    def pin_memory(self):
+        return self
+
+def collate_wrapper_random(batch):
+    return BatchLoaderRandom(batch)

modules/textual_inversion/textual_inversion.py

@@ -64,7 +64,8 @@ class EmbeddingDatabase:
 
         self.word_embeddings[embedding.name] = embedding
 
-        ids = model.cond_stage_model.tokenizer([embedding.name], add_special_tokens=False)['input_ids'][0]
+        # TODO changing between clip and open clip changes tokenization, which will cause embeddings to stop working
+        ids = model.cond_stage_model.tokenize([embedding.name])[0]
 
         first_id = ids[0]
         if first_id not in self.ids_lookup:
@@ -155,13 +156,11 @@ class EmbeddingDatabase:
 
 def create_embedding(name, num_vectors_per_token, overwrite_old, init_text='*'):
     cond_model = shared.sd_model.cond_stage_model
-    embedding_layer = cond_model.wrapped.transformer.text_model.embeddings
 
     with devices.autocast():
         cond_model([""])  # will send cond model to GPU if lowvram/medvram is active
 
-    ids = cond_model.tokenizer(init_text, max_length=num_vectors_per_token, return_tensors="pt", add_special_tokens=False)["input_ids"]
-    embedded = embedding_layer.token_embedding.wrapped(ids.to(devices.device)).squeeze(0)
+    embedded = cond_model.encode_embedding_init_text(init_text, num_vectors_per_token)
     vec = torch.zeros((num_vectors_per_token, embedded.shape[1]), device=devices.device)
 
     for i in range(num_vectors_per_token):
@@ -184,7 +183,7 @@ def write_loss(log_directory, filename, step, epoch_len, values):
     if shared.opts.training_write_csv_every == 0:
         return
 
-    if (step + 1) % shared.opts.training_write_csv_every != 0:
+    if step % shared.opts.training_write_csv_every != 0:
         return
     write_csv_header = False if os.path.exists(os.path.join(log_directory, filename)) else True
 
@@ -194,21 +193,23 @@ def write_loss(log_directory, filename, step, epoch_len, values):
         if write_csv_header:
             csv_writer.writeheader()
 
-        epoch = step // epoch_len
-        epoch_step = step % epoch_len 
+        epoch = (step - 1) // epoch_len
+        epoch_step = (step - 1) % epoch_len 
 
         csv_writer.writerow({
-            "step": step + 1,
+            "step": step,
             "epoch": epoch,
-            "epoch_step": epoch_step + 1,
+            "epoch_step": epoch_step,
             **values,
         })
 
-def validate_train_inputs(model_name, learn_rate, batch_size, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"):
+def validate_train_inputs(model_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_model_every, create_image_every, log_directory, name="embedding"):
     assert model_name, f"{name} not selected"
     assert learn_rate, "Learning rate is empty or 0"
     assert isinstance(batch_size, int), "Batch size must be integer"
     assert batch_size > 0, "Batch size must be positive"
+    assert isinstance(gradient_step, int), "Gradient accumulation step must be integer"
+    assert gradient_step > 0, "Gradient accumulation step must be positive"
     assert data_root, "Dataset directory is empty"
     assert os.path.isdir(data_root), "Dataset directory doesn't exist"
     assert os.listdir(data_root), "Dataset directory is empty"
@@ -224,10 +225,10 @@ def validate_train_inputs(model_name, learn_rate, batch_size, data_root, templat
     if save_model_every or create_image_every:
         assert log_directory, "Log directory is empty"
 
-def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
+def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
     save_embedding_every = save_embedding_every or 0
     create_image_every = create_image_every or 0
-    validate_train_inputs(embedding_name, learn_rate, batch_size, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding")
+    validate_train_inputs(embedding_name, learn_rate, batch_size, gradient_step, data_root, template_file, steps, save_embedding_every, create_image_every, log_directory, name="embedding")
 
     shared.state.textinfo = "Initializing textual inversion training..."
     shared.state.job_count = steps
@@ -255,161 +256,200 @@ def train_embedding(embedding_name, learn_rate, batch_size, data_root, log_direc
     else:
         images_embeds_dir = None
 
-    cond_model = shared.sd_model.cond_stage_model
-
     hijack = sd_hijack.model_hijack
 
     embedding = hijack.embedding_db.word_embeddings[embedding_name]
     checkpoint = sd_models.select_checkpoint()
 
-    ititial_step = embedding.step or 0
-    if ititial_step >= steps:
+    initial_step = embedding.step or 0
+    if initial_step >= steps:
         shared.state.textinfo = f"Model has already been trained beyond specified max steps"
         return embedding, filename
+    scheduler = LearnRateScheduler(learn_rate, steps, initial_step)
 
-    scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
-
-    # dataset loading may take a while, so input validations and early returns should be done before this
+   # dataset loading may take a while, so input validations and early returns should be done before this
     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
-    with torch.autocast("cuda"):
-        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file, batch_size=batch_size)
+    
+    pin_memory = shared.opts.pin_memory
+    
+    ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, cond_model=shared.sd_model.cond_stage_model, device=devices.device, template_file=template_file, batch_size=batch_size, gradient_step=gradient_step, shuffle_tags=shuffle_tags, tag_drop_out=tag_drop_out, latent_sampling_method=latent_sampling_method)
+
+    latent_sampling_method = ds.latent_sampling_method
+
+    dl = modules.textual_inversion.dataset.PersonalizedDataLoader(ds, latent_sampling_method=latent_sampling_method, batch_size=ds.batch_size, pin_memory=pin_memory)
+
     if unload:
         shared.sd_model.first_stage_model.to(devices.cpu)
 
     embedding.vec.requires_grad = True
-    optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)
+    optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate, weight_decay=0.0)
+    scaler = torch.cuda.amp.GradScaler()
 
-    losses = torch.zeros((32,))
+    batch_size = ds.batch_size
+    gradient_step = ds.gradient_step
+    # n steps = batch_size * gradient_step * n image processed
+    steps_per_epoch = len(ds) // batch_size // gradient_step
+    max_steps_per_epoch = len(ds) // batch_size - (len(ds) // batch_size) % gradient_step
+    loss_step = 0
+    _loss_step = 0 #internal
 
+    
     last_saved_file = "<none>"
     last_saved_image = "<none>"
     forced_filename = "<none>"
     embedding_yet_to_be_embedded = False
+    
+    pbar = tqdm.tqdm(total=steps - initial_step)
+    try:
+        for i in range((steps-initial_step) * gradient_step):
+            if scheduler.finished:
+                break
+            if shared.state.interrupted:
+                break
+            for j, batch in enumerate(dl):
+                # works as a drop_last=True for gradient accumulation
+                if j == max_steps_per_epoch:
+                    break
+                scheduler.apply(optimizer, embedding.step)
+                if scheduler.finished:
+                    break
+                if shared.state.interrupted:
+                    break
 
-    pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
-    for i, entries in pbar:
-        embedding.step = i + ititial_step
+                with torch.autocast("cuda"):
+                    # c = stack_conds(batch.cond).to(devices.device)
+                    # mask = torch.tensor(batch.emb_index).to(devices.device, non_blocking=pin_memory)
+                    # print(mask)
+                    # c[:, 1:1+embedding.vec.shape[0]] = embedding.vec.to(devices.device, non_blocking=pin_memory)
+                    x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
+                    c = shared.sd_model.cond_stage_model(batch.cond_text)
+                    loss = shared.sd_model(x, c)[0] / gradient_step
+                    del x
+                    
+                    _loss_step += loss.item()
+                scaler.scale(loss).backward()
+                
+                # go back until we reach gradient accumulation steps
+                if (j + 1) % gradient_step != 0:
+                    continue
+                scaler.step(optimizer)
+                scaler.update()
+                embedding.step += 1
+                pbar.update()
+                optimizer.zero_grad(set_to_none=True)
+                loss_step = _loss_step
+                _loss_step = 0
 
-        scheduler.apply(optimizer, embedding.step)
-        if scheduler.finished:
-            break
+                steps_done = embedding.step + 1
 
-        if shared.state.interrupted:
-            break
+                epoch_num = embedding.step // steps_per_epoch
+                epoch_step = embedding.step % steps_per_epoch
 
-        with torch.autocast("cuda"):
-            c = cond_model([entry.cond_text for entry in entries])
-            x = torch.stack([entry.latent for entry in entries]).to(devices.device)
-            loss = shared.sd_model(x, c)[0]
-            del x
+                pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{steps_per_epoch}]loss: {loss_step:.7f}")
+                if embedding_dir is not None and steps_done % save_embedding_every == 0:
+                    # Before saving, change name to match current checkpoint.
+                    embedding_name_every = f'{embedding_name}-{steps_done}'
+                    last_saved_file = os.path.join(embedding_dir, f'{embedding_name_every}.pt')
+                    #if shared.opts.save_optimizer_state:
+                        #embedding.optimizer_state_dict = optimizer.state_dict()
+                    save_embedding(embedding, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True)
+                    embedding_yet_to_be_embedded = True
 
-            losses[embedding.step % losses.shape[0]] = loss.item()
+                write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, steps_per_epoch, {
+                    "loss": f"{loss_step:.7f}",
+                    "learn_rate": scheduler.learn_rate
+                })
 
-            optimizer.zero_grad()
-            loss.backward()
-            optimizer.step()
+                if images_dir is not None and steps_done % create_image_every == 0:
+                    forced_filename = f'{embedding_name}-{steps_done}'
+                    last_saved_image = os.path.join(images_dir, forced_filename)
 
-        steps_done = embedding.step + 1
+                    shared.sd_model.first_stage_model.to(devices.device)
 
-        epoch_num = embedding.step // len(ds)
-        epoch_step = embedding.step % len(ds)
+                    p = processing.StableDiffusionProcessingTxt2Img(
+                        sd_model=shared.sd_model,
+                        do_not_save_grid=True,
+                        do_not_save_samples=True,
+                        do_not_reload_embeddings=True,
+                    )
 
-        pbar.set_description(f"[Epoch {epoch_num}: {epoch_step+1}/{len(ds)}]loss: {losses.mean():.7f}")
+                    if preview_from_txt2img:
+                        p.prompt = preview_prompt
+                        p.negative_prompt = preview_negative_prompt
+                        p.steps = preview_steps
+                        p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
+                        p.cfg_scale = preview_cfg_scale
+                        p.seed = preview_seed
+                        p.width = preview_width
+                        p.height = preview_height
+                    else:
+                        p.prompt = batch.cond_text[0]
+                        p.steps = 20
+                        p.width = training_width
+                        p.height = training_height
 
-        if embedding_dir is not None and steps_done % save_embedding_every == 0:
-            # Before saving, change name to match current checkpoint.
-            embedding_name_every = f'{embedding_name}-{steps_done}'
-            last_saved_file = os.path.join(embedding_dir, f'{embedding_name_every}.pt')
-            save_embedding(embedding, checkpoint, embedding_name_every, last_saved_file, remove_cached_checksum=True)
-            embedding_yet_to_be_embedded = True
+                    preview_text = p.prompt
 
-        write_loss(log_directory, "textual_inversion_loss.csv", embedding.step, len(ds), {
-            "loss": f"{losses.mean():.7f}",
-            "learn_rate": scheduler.learn_rate
-        })
+                    processed = processing.process_images(p)
+                    image = processed.images[0] if len(processed.images) > 0 else None
 
-        if images_dir is not None and steps_done % create_image_every == 0:
-            forced_filename = f'{embedding_name}-{steps_done}'
-            last_saved_image = os.path.join(images_dir, forced_filename)
+                    if unload:
+                        shared.sd_model.first_stage_model.to(devices.cpu)
 
-            shared.sd_model.first_stage_model.to(devices.device)
+                    if image is not None:
+                        shared.state.current_image = image
+                        last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
+                        last_saved_image += f", prompt: {preview_text}"
 
-            p = processing.StableDiffusionProcessingTxt2Img(
-                sd_model=shared.sd_model,
-                do_not_save_grid=True,
-                do_not_save_samples=True,
-                do_not_reload_embeddings=True,
-            )
+                    if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded:
 
-            if preview_from_txt2img:
-                p.prompt = preview_prompt
-                p.negative_prompt = preview_negative_prompt
-                p.steps = preview_steps
-                p.sampler_name = sd_samplers.samplers[preview_sampler_index].name
-                p.cfg_scale = preview_cfg_scale
-                p.seed = preview_seed
-                p.width = preview_width
-                p.height = preview_height
-            else:
-                p.prompt = entries[0].cond_text
-                p.steps = 20
-                p.width = training_width
-                p.height = training_height
+                        last_saved_image_chunks = os.path.join(images_embeds_dir, f'{embedding_name}-{steps_done}.png')
 
-            preview_text = p.prompt
+                        info = PngImagePlugin.PngInfo()
+                        data = torch.load(last_saved_file)
+                        info.add_text("sd-ti-embedding", embedding_to_b64(data))
 
-            processed = processing.process_images(p)
-            image = processed.images[0]
+                        title = "<{}>".format(data.get('name', '???'))
 
-            if unload:
-                shared.sd_model.first_stage_model.to(devices.cpu)
+                        try:
+                            vectorSize = list(data['string_to_param'].values())[0].shape[0]
+                        except Exception as e:
+                            vectorSize = '?'
 
-            shared.state.current_image = image
+                        checkpoint = sd_models.select_checkpoint()
+                        footer_left = checkpoint.model_name
+                        footer_mid = '[{}]'.format(checkpoint.hash)
+                        footer_right = '{}v {}s'.format(vectorSize, steps_done)
 
-            if save_image_with_stored_embedding and os.path.exists(last_saved_file) and embedding_yet_to_be_embedded:
+                        captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right)
+                        captioned_image = insert_image_data_embed(captioned_image, data)
 
-                last_saved_image_chunks = os.path.join(images_embeds_dir, f'{embedding_name}-{steps_done}.png')
+                        captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info)
+                        embedding_yet_to_be_embedded = False
 
-                info = PngImagePlugin.PngInfo()
-                data = torch.load(last_saved_file)
-                info.add_text("sd-ti-embedding", embedding_to_b64(data))
+                    last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
+                    last_saved_image += f", prompt: {preview_text}"
 
-                title = "<{}>".format(data.get('name', '???'))
+                shared.state.job_no = embedding.step
 
-                try:
-                    vectorSize = list(data['string_to_param'].values())[0].shape[0]
-                except Exception as e:
-                    vectorSize = '?'
-
-                checkpoint = sd_models.select_checkpoint()
-                footer_left = checkpoint.model_name
-                footer_mid = '[{}]'.format(checkpoint.hash)
-                footer_right = '{}v {}s'.format(vectorSize, steps_done)
-
-                captioned_image = caption_image_overlay(image, title, footer_left, footer_mid, footer_right)
-                captioned_image = insert_image_data_embed(captioned_image, data)
-
-                captioned_image.save(last_saved_image_chunks, "PNG", pnginfo=info)
-                embedding_yet_to_be_embedded = False
-
-            last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
-            last_saved_image += f", prompt: {preview_text}"
-
-        shared.state.job_no = embedding.step
-
-        shared.state.textinfo = f"""
+                shared.state.textinfo = f"""
 <p>
-Loss: {losses.mean():.7f}<br/>
-Step: {embedding.step}<br/>
-Last prompt: {html.escape(entries[0].cond_text)}<br/>
+Loss: {loss_step:.7f}<br/>
+Step: {steps_done}<br/>
+Last prompt: {html.escape(batch.cond_text[0])}<br/>
 Last saved embedding: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
 </p>
 """
-
-    filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')
-    save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True)
-    shared.sd_model.first_stage_model.to(devices.device)
+        filename = os.path.join(shared.cmd_opts.embeddings_dir, f'{embedding_name}.pt')
+        save_embedding(embedding, checkpoint, embedding_name, filename, remove_cached_checksum=True)
+    except Exception:
+        print(traceback.format_exc(), file=sys.stderr)
+        pass
+    finally:
+        pbar.leave = False
+        pbar.close()
+        shared.sd_model.first_stage_model.to(devices.device)
 
     return embedding, filename
 

modules/ui.py

@@ -157,22 +157,6 @@ def save_files(js_data, images, do_make_zip, index):
 
     return gr.File.update(value=fullfns, visible=True), '', '', plaintext_to_html(f"Saved: {filenames[0]}")
 
-def save_pil_to_file(pil_image, dir=None):
-    use_metadata = False
-    metadata = PngImagePlugin.PngInfo()
-    for key, value in pil_image.info.items():
-        if isinstance(key, str) and isinstance(value, str):
-            metadata.add_text(key, value)
-            use_metadata = True
-
-    file_obj = tempfile.NamedTemporaryFile(delete=False, suffix=".png", dir=dir)
-    pil_image.save(file_obj, pnginfo=(metadata if use_metadata else None))
-    return file_obj
-
-
-# override save to file function so that it also writes PNG info
-gr.processing_utils.save_pil_to_file = save_pil_to_file
-
 
 def wrap_gradio_call(func, extra_outputs=None, add_stats=False):
     def f(*args, extra_outputs_array=extra_outputs, **kwargs):
@@ -478,9 +462,7 @@ def create_toprow(is_img2img):
         if is_img2img:
             with gr.Column(scale=1, elem_id="interrogate_col"):
                 button_interrogate = gr.Button('Interrogate\nCLIP', elem_id="interrogate")
-
-                if cmd_opts.deepdanbooru:
-                    button_deepbooru = gr.Button('Interrogate\nDeepBooru', elem_id="deepbooru")
+                button_deepbooru = gr.Button('Interrogate\nDeepBooru', elem_id="deepbooru")
 
         with gr.Column(scale=1):
             with gr.Row():
@@ -1004,11 +986,10 @@ def create_ui(wrap_gradio_gpu_call):
                 outputs=[img2img_prompt],
             )
 
-            if cmd_opts.deepdanbooru:
-                img2img_deepbooru.click(
-                    fn=interrogate_deepbooru,
-                    inputs=[init_img],
-                    outputs=[img2img_prompt],
+            img2img_deepbooru.click(
+                fn=interrogate_deepbooru,
+                inputs=[init_img],
+                outputs=[img2img_prompt],
             )
 
 
@@ -1213,7 +1194,7 @@ def create_ui(wrap_gradio_gpu_call):
 
                 with gr.Tab(label="Create hypernetwork"):
                     new_hypernetwork_name = gr.Textbox(label="Name")
-                    new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"])
+                    new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "1024", "320", "640", "1280"])
                     new_hypernetwork_layer_structure = gr.Textbox("1, 2, 1", label="Enter hypernetwork layer structure", placeholder="1st and last digit must be 1. ex:'1, 2, 1'")
                     new_hypernetwork_activation_func = gr.Dropdown(value="linear", label="Select activation function of hypernetwork. Recommended : Swish / Linear(none)", choices=modules.hypernetworks.ui.keys)
                     new_hypernetwork_initialization_option = gr.Dropdown(value = "Normal", label="Select Layer weights initialization. Recommended: Kaiming for relu-like, Xavier for sigmoid-like, Normal otherwise", choices=["Normal", "KaimingUniform", "KaimingNormal", "XavierUniform", "XavierNormal"])
@@ -1259,7 +1240,7 @@ def create_ui(wrap_gradio_gpu_call):
                         with gr.Column():
                             with gr.Row():
                                 interrupt_preprocessing = gr.Button("Interrupt")
-                                run_preprocess = gr.Button(value="Preprocess", variant='primary')
+                            run_preprocess = gr.Button(value="Preprocess", variant='primary')
 
                     process_split.change(
                         fn=lambda show: gr_show(show),
@@ -1286,6 +1267,7 @@ def create_ui(wrap_gradio_gpu_call):
                         hypernetwork_learn_rate = gr.Textbox(label='Hypernetwork Learning rate', placeholder="Hypernetwork Learning rate", value="0.00001")
 
                     batch_size = gr.Number(label='Batch size', value=1, precision=0)
+                    gradient_step = gr.Number(label='Gradient accumulation steps', value=1, precision=0)
                     dataset_directory = gr.Textbox(label='Dataset directory', placeholder="Path to directory with input images")
                     log_directory = gr.Textbox(label='Log directory', placeholder="Path to directory where to write outputs", value="textual_inversion")
                     template_file = gr.Textbox(label='Prompt template file', value=os.path.join(script_path, "textual_inversion_templates", "style_filewords.txt"))
@@ -1296,6 +1278,11 @@ def create_ui(wrap_gradio_gpu_call):
                     save_embedding_every = gr.Number(label='Save a copy of embedding to log directory every N steps, 0 to disable', value=500, precision=0)
                     save_image_with_stored_embedding = gr.Checkbox(label='Save images with embedding in PNG chunks', value=True)
                     preview_from_txt2img = gr.Checkbox(label='Read parameters (prompt, etc...) from txt2img tab when making previews', value=False)
+                    with gr.Row():
+                        shuffle_tags = gr.Checkbox(label="Shuffle tags by ',' when creating prompts.", value=False)
+                        tag_drop_out = gr.Slider(minimum=0, maximum=1, step=0.1, label="Drop out tags when creating prompts.", value=0)
+                    with gr.Row():
+                        latent_sampling_method = gr.Radio(label='Choose latent sampling method', value="once", choices=['once', 'deterministic', 'random'])
 
                     with gr.Row():
                         interrupt_training = gr.Button(value="Interrupt")
@@ -1384,11 +1371,15 @@ def create_ui(wrap_gradio_gpu_call):
                 train_embedding_name,
                 embedding_learn_rate,
                 batch_size,
+                gradient_step,
                 dataset_directory,
                 log_directory,
                 training_width,
                 training_height,
                 steps,
+                shuffle_tags,
+                tag_drop_out,
+                latent_sampling_method,
                 create_image_every,
                 save_embedding_every,
                 template_file,
@@ -1409,11 +1400,15 @@ def create_ui(wrap_gradio_gpu_call):
                 train_hypernetwork_name,
                 hypernetwork_learn_rate,
                 batch_size,
+                gradient_step,
                 dataset_directory,
                 log_directory,
                 training_width,
                 training_height,
                 steps,
+                shuffle_tags,
+                tag_drop_out,
+                latent_sampling_method,
                 create_image_every,
                 save_embedding_every,
                 template_file,

modules/ui_tempdir.py

@@ -0,0 +1,62 @@
+import os
+import tempfile
+from collections import namedtuple
+
+import gradio as gr
+
+from PIL import PngImagePlugin
+
+from modules import shared
+
+
+Savedfile = namedtuple("Savedfile", ["name"])
+
+
+def save_pil_to_file(pil_image, dir=None):
+    already_saved_as = getattr(pil_image, 'already_saved_as', None)
+    if already_saved_as:
+        shared.demo.temp_dirs = shared.demo.temp_dirs | {os.path.abspath(os.path.dirname(already_saved_as))}
+        file_obj = Savedfile(already_saved_as)
+        return file_obj
+
+    if shared.opts.temp_dir != "":
+        dir = shared.opts.temp_dir
+
+    use_metadata = False
+    metadata = PngImagePlugin.PngInfo()
+    for key, value in pil_image.info.items():
+        if isinstance(key, str) and isinstance(value, str):
+            metadata.add_text(key, value)
+            use_metadata = True
+
+    file_obj = tempfile.NamedTemporaryFile(delete=False, suffix=".png", dir=dir)
+    pil_image.save(file_obj, pnginfo=(metadata if use_metadata else None))
+    return file_obj
+
+
+# override save to file function so that it also writes PNG info
+gr.processing_utils.save_pil_to_file = save_pil_to_file
+
+
+def on_tmpdir_changed():
+    if shared.opts.temp_dir == "" or shared.demo is None:
+        return
+
+    os.makedirs(shared.opts.temp_dir, exist_ok=True)
+
+    shared.demo.temp_dirs = shared.demo.temp_dirs | {os.path.abspath(shared.opts.temp_dir)}
+
+
+def cleanup_tmpdr():
+    temp_dir = shared.opts.temp_dir
+    if temp_dir == "" or not os.path.isdir(temp_dir):
+        return
+
+    for root, dirs, files in os.walk(temp_dir, topdown=False):
+        for name in files:
+            _, extension = os.path.splitext(name)
+            if extension != ".png":
+                continue
+
+            filename = os.path.join(root, name)
+            os.remove(filename)

v1-inference.yaml

@@ -0,0 +1,70 @@
+model:
+  base_learning_rate: 1.0e-04
+  target: ldm.models.diffusion.ddpm.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: "jpg"
+    cond_stage_key: "txt"
+    image_size: 64
+    channels: 4
+    cond_stage_trainable: false   # Note: different from the one we trained before
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_factor: 0.18215
+    use_ema: False
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [ 10000 ]
+        cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
+        f_start: [ 1.e-6 ]
+        f_max: [ 1. ]
+        f_min: [ 1. ]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32 # unused
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions: [ 4, 2, 1 ]
+        num_res_blocks: 2
+        channel_mult: [ 1, 2, 4, 4 ]
+        num_heads: 8
+        use_spatial_transformer: True
+        transformer_depth: 1
+        context_dim: 768
+        use_checkpoint: True
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 256
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

webui.py

@@ -10,7 +10,7 @@ from fastapi.middleware.gzip import GZipMiddleware
 
 from modules.paths import script_path
 
-from modules import devices, sd_samplers, upscaler, extensions, localization
+from modules import shared, devices, sd_samplers, upscaler, extensions, localization, ui_tempdir
 import modules.codeformer_model as codeformer
 import modules.extras
 import modules.face_restoration
@@ -23,7 +23,6 @@ import modules.scripts
 import modules.sd_hijack
 import modules.sd_models
 import modules.sd_vae
-import modules.shared as shared
 import modules.txt2img
 import modules.script_callbacks
 
@@ -32,12 +31,14 @@ from modules import modelloader
 from modules.shared import cmd_opts
 import modules.hypernetworks.hypernetwork
 
+
 queue_lock = threading.Lock()
 if cmd_opts.server_name:
     server_name = cmd_opts.server_name
 else:
     server_name = "0.0.0.0" if cmd_opts.listen else None
 
+
 def wrap_queued_call(func):
     def f(*args, **kwargs):
         with queue_lock:
@@ -86,8 +87,9 @@ def initialize():
     shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: modules.sd_models.reload_model_weights()))
     shared.opts.onchange("sd_vae", wrap_queued_call(lambda: modules.sd_vae.reload_vae_weights()), call=False)
     shared.opts.onchange("sd_vae_as_default", wrap_queued_call(lambda: modules.sd_vae.reload_vae_weights()), call=False)
-    shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: modules.hypernetworks.hypernetwork.load_hypernetwork(shared.opts.sd_hypernetwork)))
+    shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: shared.reload_hypernetworks()))
     shared.opts.onchange("sd_hypernetwork_strength", modules.hypernetworks.hypernetwork.apply_strength)
+    shared.opts.onchange("temp_dir", ui_tempdir.on_tmpdir_changed)
 
     if cmd_opts.tls_keyfile is not None and cmd_opts.tls_keyfile is not None:
 
@@ -150,9 +152,12 @@ def webui():
     initialize()
 
     while 1:
-        demo = modules.ui.create_ui(wrap_gradio_gpu_call=wrap_gradio_gpu_call)
+        if shared.opts.clean_temp_dir_at_start:
+            ui_tempdir.cleanup_tmpdr()
 
-        app, local_url, share_url = demo.launch(
+        shared.demo = modules.ui.create_ui(wrap_gradio_gpu_call=wrap_gradio_gpu_call)
+
+        app, local_url, share_url = shared.demo.launch(
             share=cmd_opts.share,
             server_name=server_name,
             server_port=cmd_opts.port,
@@ -179,9 +184,9 @@ def webui():
         if launch_api:
             create_api(app)
 
-        modules.script_callbacks.app_started_callback(demo, app)
+        modules.script_callbacks.app_started_callback(shared.demo, app)
 
-        wait_on_server(demo)
+        wait_on_server(shared.demo)
 
         sd_samplers.set_samplers()