Pedro Cuenca

Latent upscaler

3

WIP. The upscaler model does not predict the noise, but the denoised image. Therefore it needs to use schedulers that support `predict_epsilon=False`. In addition, [the noise schedule is configured from...

### Describe the bug `trained_betas` is a numpy array, and it cannot be serialized to JSON. This makes it currently impossible to save the configuration of a custom scheduler that...

Add experimental Heun scheduler

1

Without duplicating the sigmas and timesteps (reference: #1336) ➕ No duplication of sigmas and timesteps. Does not require ugly function `index_for_timestep` (or confusing oneliner repeated twice in the code) ➖...

Do not use torch.long in mps

1

Fixes #1056. Another option is to unconditionally use `torch.float32` in all platforms (both `int` and `float` are accepted as inputs), what do you think?

This should be convenient and time-saving for many users.

Restricted to order 2, no dynamic thresholding. It achieves similar quality to PNDM in 25 steps (~10s in my Mac).

Otherwise the app may crash when attempting to use a random Int as a seed.

The wheels will have PyTorch pinned to the latest stable version: https://github.com/facebookresearch/xformers/issues/533#issuecomment-1355162407 For older versions, we can recommend to install from source.

MPS: models require an initial pass for reproducibility

10

I have no idea what's causing this, I suspect something related to RNGs, as we know they don't work properly in `mps` yet. This is a pending task from our...

Dreambooth docs: minor fixes

2

Add missing section headers. README links to docs section about running inference using a saved checkpoint.