This is the conversion of SmoothMix WAN 2.2 V2.0 - https://civitai.com/models/1995784/smooth-mix-wan-22-i2vt2v-14b
I'm experimentei running the script https://github.com/Kickbub/Dequant-FP8-ComfyUI/blob/main/dequantize_fp8v2.py to turn it to FP16 to check if there is an imporvement on quality and size
There is the full FP16 conversion, around 40Gb, and some GGUF.
Please provide feedback so I know if I should run this script always before GGUF other models.
Videos made with UMT GGUF Q8 and VAE FP32
| . | F8 | F16 |
|---|---|---|
| Full | ||
| Q8 |
There is another example in the example folder
===================================================================================
if you would like to help me, it seems that runpod has a Refer thing - https://runpod.io?ref=d2452mau
| You get | I get |
|---|---|
| - A one-time credit of $5 when they sign up with your link and adds $10 for the first time - Instant access to Runpod's GPU resources |
- A one-time credit of $5 when a user signs up with your link and adds $10 for the first time - Credits on referred user spend during their first 6 months. (5% Serverless and 3% Pods) |