Sign In

Workflow, finally

Loading Images

UltraHD 3 pass hunyuan video workflow by TedBerlim.

https://civitai.com/posts/11192734

I'll hide this here under a video post, because it's very "special" combination of LORAs and configs. So not really applicable for everyone.

Just for you Nerds. ;-)

Working locally on my 4090 with 113/129 frames. (depending on resolution and passes)

=========>>> I enabled NVidia system-memory-fallback policy in Nvidia's systemcontrol, I think that's the reason for extended lenght?

----->>> you need SageAttention with Triton installed!!!! <<<<------- see post for windows installation. linux should be much easier

https://www.reddit.com/r/StableDiffusion/comments/1h7hunp/how_to_run_hunyuanvideo_on_a_single_24gb_vram_card/

So all credits goes to the following guys. i just mixed them up all together:

GITHUB:

=======

https://github.com/Tencent/HunyuanVideo

https://github.com/kijai/ComfyUI-HunyuanVideoWrapper <- get fastfp8 safetensor from kijai's huggingface space

https://github.com/zer0int/ComfyUI-HunyuanVideo-Nyan <- for long-clip SAE and shifting factors

https://github.com/NUS-HPC-AI-Lab/Enhance-A-Video <- included in Kijai's wrapper

Civit.AI Posts:

==============

Hunyuan 2step t2v and upscale <- for HiResUpscale, 2step should be enought for the most, but somehow i wanted to try 3 passes.

https://civitai.com/models/1092466?modelVersionId=1227014

Female - Face Portraits - Detailed Skin - Hunyuan <- enables sharpness and realism, not only face closeups,

https://civitai.com/models/1084718/female-face-portraits-detailed-skin-hunyuan?modelVersionId=1218022

also thxs to AIVideoEnthusiast for his prompt (in my note under prompt for more detailed control) and findings on sharpness and realism for this LORA

https://civitai.com/images/49072833

Natalie Portman Hunyuan Video Lora

https://civitai.com/models/1090839?modelVersionId=1225118

and of course also Felldude for "hosting" the search results under his fp8 checkpoint. i use instead the fastfp8 with much oversized steps, but couldn't find a basemodel with fastfp8 on Civit.AI

https://civitai.com/models/1018217?modelVersionId=1177017

===================================================================================================================================

Runtime comments for my generations on my 4090 24GB VRAM / 64GB System RAM on Windows 10!

=========>>> i enabled NVidia system-memory-fallback policy in Nvidia's systemcontrol, I think that's the reason for extended lenght?

===================================================================================================================================

1. pass duration should be around 04:30 (mm:ss)

-------

got prompt

Input (height, width, video_length) = (384, 576, 113)

Swapping 20 double blocks and 40 single blocks

Sampling 113 frames in 29 latents at 576x384 with 50 inference steps

10%|██████████▉ | 5/50 [00:27<04:06, 5.47s/it]

2. pass duration should be around 07:30 (mm:ss)

-------

Sampling 113 frames in 29 latents at 864x576 with 30 inference steps

50%|██████████████████████████████████████████████████████ | 15/30 [03:44<03:42, 14.83s/it]

3. pass duration should be around 27:30 (mm:ss) ----- if you really need that. the clothes-movements are more fluid in the 2.pass

-------

Sampling 113 frames in 29 latents at 1296x864 with 30 inference steps

7%|███████▎ | 2/30 [01:46<24:57, 53.47s/it]

OVERALL:

========

The magic lies in the combination of all the individual components. But especially in my opinion the "LORA Female - Face Portraits - Detailed Skin". With a coordinated weighting some sort of UNLOCK comes into place. Also with the "pee" LORA, but thats another section ;-)

Ohhhh. and the long clip AND customnode from zer0int (https://github.com/zer0int/ComfyUI-HunyuanVideo-Nyan) gaves me another kick.

But as I said, the combination is the magic..........

Ted

Comments