Not finding these on civitia and I need for comfy in runpod - wan2.1_t2v_14B_fp8_scaled
fp8 scaled 14B models.
These are the models I use locally to render
wan2.1_t2v_14B_fp8_e4m3fn.safetensors - pending upload
14.3 GB
LFS
Simple fp8
Runpod: (One Click deploy - ComfyUI Wan14B t2v i2v v2v) not mine.. but works good.
navigate to diffusion_models
run: download.py --model 1666198 (hope it works)
wan2.1_t2v_14B_fp8_scaled.safetensors - uploaded 4.15.25
14.3 GB
LFS
fp8 scaled 14B models.
https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
Description
FAQ
Comments (11)
You can use wget to download any download link to runpod. Including huggingface links.
ill look into that, I am learning runpod, and that means wasted money each try/fail.
@rickets_xxx thats on you.... do better
Is there any kind of advantage with the fp8 scaled vs the standard fp8 e4m3fn? Just tested them, doesn't seem really noticeably different
Shrug... still learning
According to comfyui it's like this:
"Note: The fp16 versions are recommended over the bf16 versions as they will give better results.
Quality rank (highest to lowest): fp16 > bf16 > fp8_scaled > fp8_e4m3fn"
See https://comfyanonymous.github.io/ComfyUI_examples/wan/?utm_source=substack&utm_medium=email
@funscripter627 I have noticed issues with the bf16 model(s) - thats why I did this mostly. going to load the fp16. instead of the e4m3fn I think.
fp8 scaled maintains about 2.5% quantization error versus 12.5% for pure e4m3fn with these models. It's really excellent! Also recommended to avoid fp8_fast as it tanks quality!
Those fucking thumbnails. Absolutely hilarious.
Can anyone kindly direct me towards a guide that explains the myriads of settings of the WAN nodes? My search skills are apparently not advanced enough. Using kijai workflows myself, but any guide would do.
The wrapper in comfy doesn't support scaled models. So how are you running these?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.