GitLab

Pytorch supports two variants of fp8:
--fp8_e4m3fn-text-enc (the one that seems to give better results)
--fp8_e5m2-text-enc

Improve diffusers model detection.

This doesn't affect how percentages behave in the frontend but breaks
things if you relied on them in the backend.

percent_to_sigma goes from 0 to 1.0 instead of 1.0 to 0 for less confusion.

Make percent 0 return an extremely large sigma and percent 1.0 return a
zero one to fix imprecision.

This should make it much easier to write custom nodes with kdiffusion type
samplers.

This was broken because of recent changes so I fixed it and moved it from
the experiments repo.

Unload models before getting weights for model patching.

Use the "lcm" sampler to sample them, you also have to use the
ModelSamplingDiscrete node to set them as lcm models to use them properly.

This allows changing the sampling parameters of the model (eps or vpred)
or set the model to use zsnr.

More generic clip model class that can be used on more types of text
encoders.

Don't apply weighting algorithm when weight is 1.0

Don't compute an empty token output when it's not needed.

Converting Nan to zero is a bad idea because it makes it hard to tell when
something went wrong.

This should make things that use sampler_cfg_function behave like before.

Added an input argument for those that want the denoised output.

This means you can calculate the x0 prediction of the model by doing:
(input - cond) for example.