For some reason, lately the results from diffusers are bad

[nitinmukesh]

So, I faced this problem a while back with Hunyuan where I quantized the models and produced the output but it was nowhere near to what I was expecting. I thought that it's due to quantization and didn't posted any issue.
I again faced the problem with Wan and thought the same and deleted it. Then I came across Diffsynth and used it (still using it) and the output was very nice. I checked and it is using quantization. I have also started using Video-X for Wan which introduced t2v, i2v, control for 1.3B and results are amazing for quantized version.
Once again I faced the same quality issue with Sana v1.5 models (all 3 of them) and ignored, thinking I am the only one having issue or I don't know how to use. Then I got notification of several threads with several other users facing similar issues.

I will post the link of few threads here

Wan-Video/Wan2.1#292
Wan-Video/Wan2.1#303
NVlabs/Sana#203

It's not that all models have this issue but the recent ones.
For e.g. LTX 0.9.1, Cog 3, Sana v1, Flux, Lumina, AuraFlow etc.... all produce good output.

If you see Sana thread it mention about Scheduler implementation. Is it to do with Scheduler and if so can't we have custom_scheduler just like custom_pipeline. I am not sure what is it related to (not a developer) so will leave with you guys to figure out and help us.

My test env is 8 + 16 + 42 for all models

[asomoza]

Hi!, to make it easier, can you post an example of this bad results? I prefer SANA since it would be easier and faster to test.

Sorry to ask you, but I need to see the bad results so I can compare them.

[nitinmukesh]

All of the information is here (images/generation data) even comparison results with original Sana and diffusers using same generation data. The topic is best referred as different information is posted by multiple users. If you want I can duplicate it here.
NVlabs/Sana#203

Direct link of 1st comment with sample images and more in replies
NVlabs/Sana#203 (comment)

[asomoza]

usually I like to have the examples contained in the issue here instead of having to look to other ones, mostly because in the future people won't have to search linked repositories to see if they are affected by the same problem.

Since you posted them here, there's no problem but I felt the need to clarify

[nitinmukesh]

Some sample output here for Sana model (using all 3 newly released models)

https://drive.google.com/file/d/1cDiONGMlp4eTHTGmDTLW-6fJ7XgIRGnY/view?usp=sharing

NOTE: All images have generation data embedded. You need metadata/ exif reader

You can also use
https://www.metadata2go.com/view-metadata

[asomoza]

DiffSynth-Studio doesn't seem to support Sana, what I mostly need is one example of what you consider a good result and another one of a bad with diffusers.

Also are you issues related to sana with human models? I did test sana sprint and found that it sometimes generated bad humans but this is the same for any other base model that's trained with safety precautions towards humans.

I'll test your images with prompt and update here.

[nitinmukesh]

I mentioned Wan for Diffsynth

I again faced the problem with Wan and thought the same and deleted it. Then I came across Diffsynth and used it (still using it) and the output was very nice. I checked and it is using quantization. I have also started using Video-X for Wan which introduced t2v, i2v, control for 1.3B and results are amazing for quantized version.

[asomoza]

I don't usually test with humans but I'll do an exception here, I know you used different settings but I have sana sprint setup for testing and "beautiful woman" is to simple of a prompt to test, so testing the other prompt with sana sprint I get these results:

these are cherry picked results though, half of them are bad (not in quality but deformed), but again, as expected with a model that is trained with safety towards humans, the beauty of SANA Sprint is that it's so fast that you can really cherry pick between 50 or more images in the same time Flux generates only one.

Also you mentioned WAN, I tested a lot WAN with diffusers, and I might be biased, but I haven't seen any better results with any other implementation than diffusers.

wan-i2v.mp4

wan-i2v_squish_49s.mp4