dg845
dg845
There's currently an issue that happens in the script as follows: if `student_timesteps` takes on the value `noise_scheduler.config.num_train_timesteps - 1 (= 999)`, `alpha_schedule[noise_scheduler.config.num_train_timesteps - 1] == 0` because `noise_scheduler` has...
If I set `alphas_schedule[-1] == 2**-24`, the student predicted original sample `student_x_0` will not have any `nan`s, but whenever `student_timesteps` is `noise_scheduler.config.num_train_timesteps - 1` the predicted x_0 will have large...
> Probably a very dumb ask but have you seen any widely different training dynamics without enabling zero-terminal SNR? I've only tested on very small toy checkpoints like [`dg845/tiny-random-stable-diffusion`](https://huggingface.co/dg845/tiny-random-stable-diffusion) due...
I think there are a few additional arguments that need to be explicitly supplied for the scripts to not raise an error. Something close to the minimal set of arguments...
Hi @SteamedGit the ADD implementation is nominally complete but I have not been able to test whether the script can distill good models (e.g. for SD v1.5) yet.
@cjt222 sorry, I haven't been able to finish testing it yet. Will hopefully find more time to work on it soon 😅.
The original [UFOGen paper](https://arxiv.org/pdf/2311.09257.pdf) doesn't explicitly give an algorithm for multistep sampling (and there is no official implementation currently available AFAIK); the current multistep sampling implementation is my current best...
Since UFOGen models typically use the same architecture and initial weights as a pretrained diffusion model (and in particular [Stable Diffusion v1.5](https://huggingface.co/runwayml/stable-diffusion-v1-5), see sections 4.2 and 5.1 of the [paper](https://arxiv.org/pdf/2311.09257.pdf)),...
Assuming the multistep sampling strategy described in https://github.com/huggingface/diffusers/pull/6133#issuecomment-1849905324 is correct, the `UFOGenScheduler` ends up being very similar to `LCMScheduler`, as both resolve the `predicted_original_sample` $`\boldsymbol{x}_0`$ and then sample from the...
Yes (although as far as I know there is no official implementation to compare the PR to, and I'm not sure if there are any publicly available UFOGen checkpoints to...