An up-to-date repo with all the necessary files can be found here: https://git.coom.tech/mrq/stable-diffusion-utils
**!**WARNING**!** **!**CAUTION**!** ***DO NOT POST THE REPO'S URL ON 4CHAN*** **!**CAUTION**!** **!**WARNING**!**
`coom.tech` is an automatic 30-day ban if posted. I am not responsible if you share that URL. Share the [rentry](https://rentry.org/sd-e621-textual-inversion/) instead.
You can also extend this into any other booru-oriented model, but you'll have to modify the pre-processing script according to the site images were pulled from. The general concepts still apply.
Below is a list of terms clarified. I notice I'll use some terms interchangably with other concepts. These do not necessarily cover everything that's generally related to Stable Diffusion, but moreso about Textual Inversion and terms I'll use that needs disambiguation:
*`Textual Inversion`: the method of "training" your embedding; comparable to training a model, but not entirely accurate.
*`training`, `learning`: running Textual Inversion to improve your embedding
*`subject`: a character / object / noun of what you're trying to train against. For e621 (or another booru) applications, it's extremely likely it's a character. Textual Inversion excels at training against subjects.
*`style`: an artist's style. Textual Inversion can also incorporate subjects in a style.
*`source content/material`: the images you're using to train against; pulled from e621 (or another booru)
*`embedding`: the trained "model" of the subject or style in question. "Model" would be wrong to call the trained output, as Textual Inversion isn't true training
I've burnt through seven or so models trying to train two of my hazubandos, each try with different methods. I've found my third attempt to have very strong results, yet I don't recall exactly what I did to get it. My later subjects failed to yield such strong results, so your mileage will greatly vary depending on the subject/style you're training against.
What works for you will differ from what works for me, but do not be discouraged if output during training looks decent, but real output in txt2img and img2img fails. Just try different, well constructed prompts, change where you place your subject, and also try and increase the size a smidge (such as 512x704, or 704x512). I've thought I've had embeddings failed, when it just took some clever tweaking for decent output.
The first step of training against a subject (or art style) is to acquire source content. Hugging Face's instructions specify having three to five images, cropped to 512x512, but there's no hard upper limit on how many, nor does having more images have any bearings on the final output size or performance. However, the more images you use, the harder it'll take for it to converge (despite convergence in typical neural network model training means overfitment).
I cannot imagine a scenario where you should stick with low image counts, such as selecting from a pool and pruning for the "best of the best". If you can get lots of images, do it. While it may appear the test outputs during training looks better with a smaller pool, when it comes to real image generation, embeddings from big image pools (140-190) yieled far better results over later embeddings trained on half the size of the first one (50-100).
If you're lacking material, the web UI's pre-processing tools to flip and split should work decently enough to cover the gap for low content. Flipping will duplicate images and flip them across the Y axis, (presumably) adding more symmetry to the final embedding, while splitting will help deal with non-square content and provide good coverage for partially generating your subject (for example, bust shots, waist below, chest only, etc.).
If you rather would have finely-crafted material, you're more than welcome to manually crop and square images. A compromise for cropping an image is to expand the canvas size to square it off, and then fill the new empty space with colors to crudely blend with the background, and crudly adding color blobs to expand limbs outside the frame. It's not that imperative to do so, but it helps.
The above tips all also apply to training a style, but some additional care needs to be taken:
***Avoid*** having a recurring subject. Textual Inversion excels at training against a recurring element, especially a subject. It's very easy for your embedding to associate with a particular character moreso than a particular style. Minimize your training material having recurring subjects.
If you already have an embedding trained for a subject, and the artist you're training against has art including that character, use that character's trained embedding. I've found it gives very promising results during training, rather than using one after the fact. It's very, very hard to get txt2img to generate an image using a subject embedding and a style embedding without having to compromise one for the other.
Use the automatic pre-processing script in the web UI to flip and split your source material, as you don't have to focus on a particular subject for training. You can get very strong results by introducing style traits that aren't tied to a specific orientation.
You are not required to actually run this, as this script is just a shortcut to manually renaming files and curating the tags, but it cuts the bulk work of it.
Included in the repo under [`./utils/renamer/`](https://git.coom.tech/mrq/stable-diffusion-utils/src/branch/master/utils/renamer) is a script for tagging images from e621 in the filename for later user in the web UI.
You can also have multiple variations of the same images, as it's useful if you're splitting an image into multiple parts. For example, the following is valid:
Clone [this repo](https://git.coom.tech/mrq/stable-diffusion-utils), open a command prompt/terminal at `./utils/renamer/`, and invoke it with `python3 preprocess.py`
I've yet to actually test this, but seeing that the prompt limit has been lifted for normal generation, I can assume it's also lifted for training. If you're feeling adventurous, you can adjust the character limit in the script to 240.
### Caveats
There's some "bugs" with the script, be it limitations with interfacing with web UI, or oversights in processing tags:
* commas are dropped both to save on filename/token count, and because the web UI will drop them anyways. This shouldn't matter so much, as commas only gives nuanced output when used, but for the strictest of users, be wary of this problem
* tags with parentheses, such as boxers_(clothing), or curt_(animal_crossing), the web UI will decide whatever it wants to when it comes to processing parentheses. I've seen some retain just the opening `(`, and some just the closing `)`, and some with multiple dangling `)`. At the absolute worst, it'll leak and emphasis some tokens during training.
* Species tags seemed to not be included in the `tags.csv`, yet they OBVIOUSLY affect the output. I haven't taken close note of it, but your results may or may not improve if you manually tag your species, either in the template or the filenames (whether the """pedantic""" reddit taxonomy term like `ursid` that e621 uses or the normal term like `bear` is prefered is unknown).
* filtering out common tags like `anthro, human, male, female`, could have negative effects with training either a subject or a style. I've definitely noticed I had to add negative terms for f\*moid parts or else my hazubando will have a cooter that I need to inpaint some cock and balls over. I've also noticed during training a style (that both has anthros and humans), a prompt associated with something anthro will generate something human. Just take notice if you don't foresee yourself ever generating a human with an anthro embedding, or anthro with a human embedding. (This also carries to ferals, but I'm sure that can be assumed)
The final piece of the puzzle is providing a decent template to train against. Under `./stable-diffusion-webui/textual_inversion_templates/` are text files for these templates. The Web UI provides rudimentary keywords (\[name\] and \[filewords\]) to help provide better crafted prompts used during training. The pre-processing script handles the \[filewords\] requirement, while \[name\] will be where you want the embedding's name to plop in the prompt.
I've had decent results with just that for training subjects. I've had mixed results with expanding that by filling in more artists to train against, for example:
would theoretically help keep the embedding from "learning" the art style itself of your subject, but again, your mileage may vary, and wouldn't use this first. I still need more tests between an embedding trained with one over the other template.
I've yet to test results when training like that, so I don't have much anecdotal advice, but only use this if you're getting output with little variation between different prompts.
Once you've managed to bang out your training template, make sure to note where you put it to reference later in the UI.
Now that everything is set up, it's time to start training. For systems with adequate enough VRAM, you're free to run the web UI with `--no-half --precision full` (whatever "adequate entails"). You'll take a very slight performance hit, but quality improves barely enough I was able to notice.
Run the Web UI, and click the `Textual Inversion` tab.
Create your embedding to train on by providing:
* a name
- can be changed later, it's just the filename, and the way to access your embedding in prompts
* the initialization text
- can be left \*
- it's only relevant for the very beginning training
- for embeds with zero training, it's effectively the same as the initialization text. For example, you can create embeds for shortcut keywords to other keywords. (The original documentation used this to """diversify""" doctors with a shortcut keyword)
* vectors per token
- this governs how much "data" can be trained to the token
- these do eat up how many tokens are left for the prompt, for example, setting this to 16 means you have 16 less tokens used for prompts
- a good range is 12 to 16, but the more you can afford the better. Given the recent change to the prompt limitation, you *could* easily just set this to 24 or 32 without worry, but I haven't personally tested the additional caveats that applies when going beyond the initial 75 tokens limit.
Click create, and the starting file will be created.
Afterwards, you can pre-process your source material further by duplicating to flip (will remove the filenames if you preprocessed them already, so beware), or split (presumably will also eat your filenames).
Next:
* select your embedding to train on in the dropdown
* if you're adventurous, adjust the learning rate. The default of `0.005` is fine enough, and shouldn't cause learning/loss problems, but if you're erring on the side of caution, you can set it to `0.0005`, but more training will be needed.
* pass in the path to the folder of your source material to train against
* put in the path to the prompt file you created earlier. if you put it in the same folder as the web UI's default prompts, just rename the filename there
* adjust how long you want the training to be done before terminating. Paperspace seems to let me do ~70000 on an A6000 before shutting down after 6 hours. An 80GB A100 will let me get shy of the full 100000 before auto-shutting down after 6 hours.
Afterwards, hit Train, and wait and watch your creation come to life.
If you didn't pre-process your images with flipped copies, I suggest midway through to pause training, then use ImageMagick's `mogrify` to flip your images with `mogrify -flop *` in the directory of your source material. I feel I've gotten nicer quality pictures because of it over an embedding I trained without it (but with a different prompt template).
## Using the Embedding
Using your newly trained embedding is as simple as putting in the name of the file in the prompt. Before, you would need to signal to the prompt parser with `<token>`, but it seems now you do not. I don't know if still using `<>` has any bearings on output, but take note you do not need it anymore.
Do not be discouraged if your initial output looks disgusting. I've found you need a nicely crafted prompt, and increasing the resolution a few notches will get something decent out of it. Play around with prompts in the thread, but I've found this one to finally give me [decent output](https://desuarchive.org/trash/thread/51397474/#51400626) (credits to [anon](https://desuarchive.org/trash/thread/51387852/#51391540) and [anon](https://desuarchive.org/trash/thread/51397474/#51397741) for letting me shamelessly steal it for my perverted kemobara needs):
And and adjusted one of the above that I found to yield very tasteful results:
```
uploaded on e621, explict content, by [Pino Daeni:__e6_artist__:0.75] and [chunie:__e6_artist__:0.75], (photography, sharp details, detailed fur, detailed eyes:1.0), <TOKEN>, hairy body, <FLAVORS>
```
where `<TOKEN>` is the name of the embedding you used, `<FLAVORS>` are additional tags you want to put in, and `__e6__artist__` is used with the Wildcards third-party script (you can manually substitute them with other artists of your choosing for subtle nuances in your ouptut).
Ordering ***really*** matters when it comes to your embedding, and additionally the weight of your embedding. Too early in the prompt, and the weight for other terms will greatly fall off, but too late in the prompt, and your embedding will lose it's influence. Too much weight applied to your embedding, and you'll deepfry your output.
If you're using an embedding primarily focused on an artstyle, and you're also using an embedding trained on a subject, take great care in your weights on your additional embedding. Too much, even the smallest amount, and you'll destroy your style's embedding in the final output.
I've mentioned adding in a drop-in replacement for `dataset.py`, with fancier stuff, like an easier way to grab tags, and to shuffle during training, but so far I don't think it's necessary. It also messes with `git pull`s, as any future updates will need intervention if that file updates.
The initial need to "fix" it was just to not use commas, but it also updated to accept booru strings.
I will try later to see if the grandeur of shuffling tags has an effect, but I imagine it's minor at most.
