Which checkpoint is the best?
@tlennon-ie which checkpoint is best for 1.1 and 1.0? the last one? I ask because sometimes, when I train, the second to last can be the best due to overfitting.
@jqlive
Version 1 had an issue with the dataset due to an AI-Toolkit bug which caused Qwen to only train on the target images. It did still have an effect but there's adverse effects that were unexpected, where it made changes outside of the skin a bit too much, such as the hair or overall textures in the background too. I left version 1 uploaded to both HF and Civitai as some users still prefer that noise artifact which kind of mimics natural grain from camera profiles.
Version 1.1 was a training run where I fixed that bug and the results are cleaner but might have less overall effect. It's not overtrained so my recommendation would be to use qwen-edit-skin_1.1_000002750.safetensors
for the full effect and at strengths 1 to 1.5 . You should see less artifacts that are unexpected but you may see less overall skin noise changes because it's focusing on the right differences between my control+target dataset rather than artificially adding bad noise everywhere