HUNYUAN | Test bench9 seconds generation time per video (*)

⚠️read disclaimer below⚠️

💥 INCREDIBLE MODEL JUST CAME OUT 💥 I'm speechless.

Kijai, our lord, has already made huge improvements to allow us to run this gold on consumer hardware with great speed. So, here we go:
With settings included in this workflow you can easily explore the latent space capabilities of this 13B beast and do all stress tests you want.

Disclaimer:

⚠️

I’ve shared this workflow, and it works perfectly fine for me and others. Confirmed working on 5 machines at the time writing this update.
However, some users have reported conflicts that cause comfy to stop functioning.

This issue could be related to recently released video nodes, Torch, or who knows what else.
Please keep this in mind before using it.

Consider this a fair warning:
I’ve done my part by sharing my work than needed to be shared with you.
DON'T SLEEP ON THIS MODEL!
Good luck!

__________________________________________________________________________________________________

around 16gb vram on lowest settings but not really sure.

⚠️To run this model you need to install torch on windows, wich is kinda pain in the ass.
Fortunately purz wrote an installation guide for all of us: link

⚠️ beware⚠️
this model is completly uncensored. It has a understanding of human anatomy at all levels.

(*) rtx3090 - vid2vid mode - 0.6 denoise - 512x320 - 17frames - 12 steps

__________________________________________________________________________________________________

__________________________________________________________________________________________________

if you want to see some real magic try this settings for txt2vid
720x480 - 25 steps - 17 frames
it takes 1:05 sec per clip on 3090 and results are .. unbelievable.