Uploaded here are some INT8 conversions of various Wan2.2 14B I2V models. All credits go to the original authors

Note: Very few BF16 originals (DaSiWa released v11 of his custom wan2.2 in int8 a bit ago) of custom wan2.2 models like those found on civitai exist, so these models were made in as a minimally lossy way possible as Claude described it.

Steps taken as follows:

Diff custom FP8 with base Wan FP8 -> Take diff and merge with base FP16 Wan -> convert FP16-ish Wan to INT8...

Base Wan FP8 and FP16 from https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged

Conversion to INT8 from https://github.com/Comfy-Org/comfy-model-tools

Quality looked 'acceptable' to me but YMMV Im guessing since most/all wan 'finetunes' are just a bunch of loras merged with base wan2.2 (there maybe some secret sauce from the author too, i dunno), claude figured this way would lose minimal quality.

If the original authors would like their models removed, i will gladly do so.

Tested with: provided workflow (or you can use your own) RTX 3060 6GB VRAM w/ 16GB RAM

With these ComfyUI flags --lowvram --windows-standalone-build --enable-manager --enable-manager-legacy-ui

8 Steps 1 CFG 512x512 upscaled to 1024x1024

Took about 12 minutes for a 15 second video using SVI2.0

Merge script also provided

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support