[Model Request]新增对MiniMax H3的视频模型支持

[Goan114]

MiniMax H3视频生成模型于近日发布。模型性能十分优秀，但是20 steps需要的计算时间真的太长......
可否新增对MiniMax H3视频模型的支持？

[KarinMiku]

indeed, this is the only thing that minimax h3 needs to be perfect! the text enconder and vae are really fast already

[Goan114]

indeed, this is the only thing that minimax h3 needs to be perfect! the text enconder and vae are really fast already

Yes, the model inference is really fast too! I have 11s/it on my RTX 5080 (16G VRAM) laptop, almost 2-3x faster than Wan 2.2's single step!
However the model still needs 20+ steps to generate, that's problem.
If they can release a 4 step lora, or inference framework that can increase inference speed, that's better!

[kingp0dd]

I'm surprised with how fast h3 is. If it can be distilled, it can match or even be faster than ltx in my case

[overallbit]

indeed, this is the only thing that minimax h3 needs to be perfect! the text enconder and vae are really fast already

Yes, the model inference is really fast too! I have 11s/it on my RTX 5080 (16G VRAM) laptop, almost 2-3x faster than Wan 2.2's single step! However the model still needs 20+ steps to generate, that's problem. If they can release a 4 step lora, or inference framework that can increase inference speed, that's better!

what resolution is that? 1366x768?

[Goan114]

indeed, this is the only thing that minimax h3 needs to be perfect! the text enconder and vae are really fast already

Yes, the model inference is really fast too! I have 11s/it on my RTX 5080 (16G VRAM) laptop, almost 2-3x faster than Wan 2.2's single step! However the model still needs 20+ steps to generate, that's problem. If they can release a 4 step lora, or inference framework that can increase inference speed, that's better!

what resolution is that? 1366x768?

864×480. Same resolution I used in WAN 2.2.

[kananaestate]

+1