<h2 id="!!!important!!!-uyowjbaps">!!!Important!!!</h2>Merging SDXL with Kolors might not be useful to transfer the SDXL knowledge to Kolors. It is more like adding random noise to Kolors which might reverse some overfitted training effect but not the way to let Kolors learns more.This is an experiment merge with (Kolors 0.8 + Openxlv3 0.2(SDXL ft)) + 0.02 perturbedAll layer merged except add_embedding.linear_1.weight due to mismatch dimAt 2024-07-08, Kolors inference is very limited. I am using this repo for the testing <a target="_blank" rel="ugc" href="https://github.com/kijai/ComfyUI-KwaiKolorsWrapper￼Backup">https://github.com/kijai/ComfyUI-KwaiKolorsWrapper Backup</a> the original Kolors model first and replace the merged verion to ComfyUI\models\diffusers\Kolors\unet\diffusion_pytorch_model.fp16.safetensors

Files with SHA256: 5ba57008830c2a7287e8ec092efd3dc64c74f0671b6eda225f4dc17bffc24414