Open narhamraatallah opened 1 week ago
supposedly, the captions are properly loaded, but once I start training I got this error:
Error join() argument must be str, bytes, or os.PathLike object, not 'NoneType' × Info Generated train script at train.bat × Error cannot identify image file 'C:\pinokio\api\fluxgym.git\datasets\368e9aa8-71ab-4421-aabe-7fd5347ea46d\0 22.txt'
text files should just fill the captions, not be considered as images
it seems like the same thing here : https://github.com/cocktailpeanut/fluxgym/issues/30 where the guy says " When you load images + text captions the matching of captions to images is done properly but the script considers all files (images + text captions) as images and training fail when trying to resize a caption text file.... "
can confirm this bug, the preparation phase recognized that we already have caption txt files for each image, but the training phase tried to run the txt files as images, hence the error. The training also mistakes as more photos, example we have 15 images+ captioning the training phase will take it like 30 photos
I made a workflow for ComfyUI, the training is abit faster (1it/s), and doesn't have any of the issues I had with FluxGym; (loads and reload the run if few seconds, properly loads .txt files, more parameters, stay in the ComfyUI environment) it's just better
can we add our custom captions in txt files ?