Here is the ComfyUI workflow pictured:
https://civarchive.com/models/1112672/fasthunyuanvideo6step
Posting two derivative HunyuanVideo models here for convenience for others. I originally found them here:
https://huggingface.co/Kijai/HunyuanVideo_comfy/tree/main
You can find the VAE there as well ^^^
Files Section has two files --->
Full model is the Regular HunyuanVideo FP8 model (CFG6, Flow 9).
Pruned model is the FastHunyuanVideo FP8 model (CFG 9, Flow 17).
N.B. I realize that they are the exact same size, but they are in fact different files for different purposes.
FastHunyuanVideo also has a LORA.
LORA: https://huggingface.co/Kijai/HunyuanVideo_comfy/blob/main/hyvideo_FastVideo_LoRA-fp8.safetensors
Description
FastHunyuanVideo Model
FAQ
Comments (18)
How much Vram is needed to run FP8? Doe it fit in 12GB ?
Try this maybe:
https://civitai.com/models/1048302?modelVersionId=1176230
But with 12gb my guess is you will be rather limited.
@az420 12 GB is enough. i use 11,6 gb.
so do I just download this and plug it in, or does it run itself
These are the main FP8 checkpoints, you'll want a HV workflow, the VAE, the text_encoders, etc... Download any of the videos you see below and drag them on ComfyUI for a preset example. You might need to configure the paths necessary for loading the various shards though...
That would have been good that the image of the workflow contain the workflow but it doesn't
good call, linked above now! https://civitai.com/models/1112672/fasthunyuanvideo6step
This workflow uses SageAttention, which might not work on Windows. Has anyone had success running this on Windows?
You can use one of the other optimizers, they are in the list :)
On my computer managed to install sageattention on Windows. And it works very well.
sagaattention works on windows 10 very well. I have an RTX 4060 ti 16 GB. Here is the link how to install it. https://old.reddit.com/r/StableDiffusion/comments/1h7hunp/how_to_run_hunyuanvideo_on_a_single_24gb_vram_card/
works on W10.
Does this gone into the Lora folder or Checkpoints?
this is a checkpoint, no lora is this big. It's the fast version of the fp8 model so it requires less steps for good generation
Needs 4x-ultrasharp.pth? how? Should I just download the model and paste in some folder?
https://civitai.com/models/116225/4x-ultrasharp
Put it in the models/ESRGAN folder (spelling might not be perfect there...)
Details
Files
hunyuanFastVideo_fasthunyuanvideo720p.safetensors
Mirrors
hunyuanFastVideo_fasthunyuanvideo720p.safetensors
Mirrors
hunyuanVideo_720CfgdistillFp8.safetensors
hunyuanVideo_720CfgdistillFp8.safetensors
hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors
hunyuanVideo_720CfgdistillFp8.safetensors
hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors
hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors
hunyuan_video_720_fp8_e4m3fn.safetensors
hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.