https://huggingface.co/datasets/StefanFalkok/ComfyUI_portable_torch_2.11.0_cu130_cp313_sageattention_triton - my new ComfyUI build with torch 2.11.0cu130 + sageattention + my workflows
Hi! I intrduce my WAN 2.2 Models 8 steps
NEW! Wan_2.2_I2V_SVI_2_PRO_8steps models
merged with SVI 2 PRO loras - you don't need choose SVI lora to make long video
https://huggingface.co/StefanFalkok/Wan_2.2_I2V_SVI_2_PRO_8steps - merged models + special workflow for those models + hunyuanvideofoley to add sound towards videos
Download
FP8/FP16 + Workflows - https://huggingface.co/StefanFalkok/Wan_2.2_10steps/tree/main
GGUF + Workflows - https://huggingface.co/StefanFalkok/Wan_2.2_10steps_GGUF/tree/main (Q8_0 and Q4_K_M)
Everything u need - just load model without light lora, set 8 steps (4/4), set 2cfg on High Noise and cfg1 on low noise, set 81 frame (16 framerate), set resolution from 576p to 720p
I merged original WAN 2.2 Models from ComfyUI repository with LightX T2V Rank 256 bf 16 (https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank256_bf16.safetensors)
fp8 results on rtx 5080
1024x576 - around 3 minutes
720p - around 5 minutes
Q8_0 and FP16 take a 20% more time to generate video, but you get more quality and stable result
My TG Channel - https://t.me/StefanFalkokAI
My TG Chat - https://t.me/+y4R5JybDZcFjMjFi
Description
Wan 2.2 I2V HighNoise 10 Steps FP8
FAQ
Comments (14)
Finally merged! Great.
Would have great success if it wasn't only for high-end 50XX GPUs. Do you plan on merging Q3_K_M / Q4_K_S for low-VRAM gen?
Hi. I think I'll quantize these models in the near future. I recommend use fp8 model and download my workflow in description at the moment
I want ask you - maybe you want Q4_K_M for better quality or u need Q4_K_S model ?
No problem on a 4080super with 32GB RAM. You're exaggerating.
I can use fp8 fine on my 3090 with 24GB VRAM.
@Jellai it's very cool
Hi.how you use f16 ? 28gb is bigger than 16gb vram?
Hi. I use Q8_0 instead because the quality is closer to FP16, but weights two times less and it eats up as many resources as FP8 model. So if you want use FP16 model, but you have a low VRAM and RAM - use Q8_0
@Stefan_Falkok .thank you. i asked chatgpt and he tootally confiused me. he told f8 is better q8. after i showed many link about that GGFU untill he accepted that GGFU can useed insted of fp.i want to sure that as you say Q8_0 is GGUF and better than FP8?
@faryarjy100 chatgpt often lies about local models. In my experience , Q8_0 is better in quality and more stable in prompts than FP8, and Q8_0 closest in quality to FP16. So I recommend use Q8_0 model, if you haven't enough VRAM/RAM
with 5080 16vram and 32 gig ram on laptop what you sugest?
@faryarjy100 oh. I thought you have 64 ram, and you'd easily used Q8_0, but I'm in doubt. You can try use FP8 and Q8_0. Otherwise, I'll release models with less quantization. But I recommend increase RAM Up to 64 (better option is 96 or 128 RAM) to use Q8_0 model
i will increase as your recommended.thank you so mach for your patience and guidance.
@faryarjy100 I'm glad to help