forked from mrq/DL-Art-School
36 lines
2.8 KiB
Markdown
36 lines
2.8 KiB
Markdown
# Working with Gaussian Diffusion models in DLAS
|
|
|
|
Diffusion Models are a method of generating structural data using a gradual de-noising process. This process allows a
|
|
simple network training regime.
|
|
|
|
This implementation of Gaussian Diffusion is largely based on the work done by OpenAI in their paper ["Diffusion Models
|
|
Beat GANs on Image Synthesis"](https://arxiv.org/pdf/2105.05233.pdf) and ["Improved Denoising Diffusion Probabilistic
|
|
Models"](https://arxiv.org/pdf/2102.09672).
|
|
|
|
OpenAI opened sourced their reference implementations [here](https://github.com/openai/guided-diffusion). The diffusion
|
|
model that DLAS trains uses the [gaussian_diffusion.py](https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/gaussian_diffusion.py)
|
|
script from that repo for training and inference with these models. We also include the UNet from that repo as a model
|
|
that can be used to train a diffusion network.
|
|
|
|
Diffusion networks can be re-purposed to pretty much any image generation task, including super-resolution. Even though
|
|
they are trained with MSE losses, they produce incredibly crisp images with FID scores competitive with the best GANs.
|
|
More importantly, it is easy to track training progress since diffusion networks use a "normal" loss.
|
|
|
|
Diffusion networks are unique in that during inference, they perform multiple forward passes to generate a single image.
|
|
During training, these networks are trained to denoise images over 4000 steps. In inference, this sample rate can be
|
|
adjusted. For the purposes of super-resolution, I have found that images sampled in 50 steps to be of very good quality.
|
|
This still means that a diffusion generator is 50x slower than generators trained in different ways.
|
|
|
|
What's more is that I have found that diffusion networks can be trained in the tiled methodology used by ESRGAN: instead
|
|
of training on whole images, you can train on tiles of larger images. At inference time, the network can be applied to
|
|
larger images than the network was initially trained on. I have found this works well on inference images within ~3x
|
|
the training size. I have not tried larger, because the size of the UNet model means that inference at ultra-high
|
|
resolutions is impossible (I run out of GPU memory).
|
|
|
|
I have provided a reference configuration for training a diffusion model in this manner. The config performs a 2x
|
|
upsampling to 256px, de-blurs it and removes JPEG artifacts. The deblurring and image repairs are done on a configurable
|
|
scale. The scale is [0,1] passed to the model as `corruption_entropy`. `1` represents a maximum correction factor.
|
|
You can try reducing this to 128px for faster training. It should work fine.
|
|
|
|
Diffusion models also have a fairly arcane inference method. To help you along, I've provided an inference configuration
|
|
that can be used with models trained in DLAS. |