- implementation of new image loading node (by me) to be able to multi-select one or more images for a run (will be executed one after another)
- dynamic prompts/wildcards now supported through Impact-Pack, syntax here - txt files have to be in the folder BEFORE comfy starts to be recognized
- introduction of refiner functionality
- switch for turning off seperate VAE to use baked-in VAE of checkpoint
- replaced the single lora loaders and their 6 activation buttons with rgthree's lora stack loader to load as many as you want (same for detailer loras)
- moved all model selection stuff into second row for more streamlined process
- removal of all group nodes & abstraction through subgraphs
- fixed v-pred model bug that didn't apply v-parameterization
- fixed modelname not saving correctly for civitai
- overhauled notes in the workflow for better understanding & improving your results
- global clipskip now happens AFTER lora loading (correct implementation, no difference to outcome quality)
wanted to add recommended detailer values you can just activate/deactivate (more vram but better results) but this is currently not possible since multiple custom_nodes still cause issues with subgraphs being bypassed