Hunyuan video not support negative prompt？

[trouble-maker007]

the official repo has supported negative prompt in inference, https://github.com/Tencent/HunyuanVideo/blob/main/hyvideo/inference.py#L505,
does diffusers has any plans to add this feature

[a-r-r-o-w]

HunyuanVideo released a guidance-distilled model so that CFG is not required for generating outputs. While it is still possible to use CFG the same way as one would with a normal model (non-guidance-distilled), it does not really produce better outputs but requires > 2x the generation time.

They also default the guidance scale to 1.0 so that CFG is not used (https://github.com/Tencent/HunyuanVideo/blob/0d44cc02aaf4a2a00b5fc10f5b0c23ee432e50a4/hyvideo/config.py#L340). From our limited experiments, the generated videos with CFG look very burnt out, oversaturated, or just bad in general. If there's a good usecase or better examples compared to the defaults we tried, we'd be super happy to support it :)

[bghira]

you have to skip the first step for CFG (or two)

[a-r-r-o-w]

@bghira If it works by skipping CFG in the first few steps and you think the results are promising/better compared to without CFG (so that it's worth 2x the time/memory), I think we could support it. Btw, we don't want to impose any restrictions by not allowing it -- it will most likely be made possible with modular diffusers so I'm not sure if it's worth putting effort into working with current pipelines

[Georgefwt]

Based on my personal experiments, setting the embedded-cfg-scale between 1-2 allows the actual cfg-scale to be adjusted to 6-7 without experiencing burnt out or oversaturation effects. Otherwise, if the embedded-cfg-scale remains at its default value of 6.0, the oversaturation appears to be quite severe.

[a-r-r-o-w]

That makes sense! Applying CFG on a guidance-distilled model with guidance_scale=A and embedded_guidance_scale=B is effectively similar to applying a guidance_scale of A*B on the non-distilled model from how the math works out

[a-r-r-o-w]

Added true CFG support in #10837 since we required it for SkyReels T2V, so the normal Hunyuan also works with it. Going forward, I think we'll stop defaulting to not having true CFG if the only model released is guidance-distilled by a team. This is to prevent confusing naming of guidance_scale meaning "true CFG" in some pipelines, while meaning "embedded CFG" in other pipelines. We will probably also have to do a breaking change at some point to fix this discrepancy cc @yiyixuxu