Edna

This may solve the following:  https://github.com/mit-han-lab/ComfyUI-nunchaku/issues/65 should be fixed by this PR (I do not use comfy ui, dunno how it manages models) I believe https://github.com/mit-han-lab/ComfyUI-nunchaku/issues/57 is caused by...

I have something really weird happening here. Feeding in a list of prompts (even if the list is only one prompt) results in a really bad video. It might be...

I checked further, and it happens when the prompt and negative prompt are length one lists (even with the list being [None], I think), but not when the prompt is...

@foreverpiano I don't believe so, as passing None in a list to negative_prompt also seems to trigger it. It also looks suspiciously like cfg baking, which I can't be certain...

@feizc In scheduler/scheduler_config.json, change `"invert_sigmas": false,` to `"invert_sigmas": true,`

@tin2tin Check the main branch docs https://huggingface.co/docs/diffusers/main/en/api/pipelines/mochi

@agneet42 You can add variant="refs/pr/18" to the model loading statement in the mean time

Hi @a-r-r-o-w! Thanks for the great reply. I've also noticed that different precisions seem to change the output drastically, and I have a theory that the model is extremely reliant...

@sayakpaul Unfortunately, I'm running into merge conflicts between layerwise upcasting and the current main branch. `mochi` got deleted, otherwise I would use it

Okay, here's what I have after some testing. ```python from diffusers import MochiPipeline, MochiTransformer3DModel from diffusers.models.transformers.transformer_mochi import MochiTransformerBlock import torch from diffusers.utils import export_to_video # The below defines how many...