Trained on ChilloutMix for 400 impressions per image (25 repeats X 4 epochs X 4 gradient accumulate steps), batch 6, Lion optimizer, 0.00005 unet LR, 0.00002 tenc LR. 207 manually-tagged images in 6 categories. This was an intermediate save from the ~26th footjob model training test, and the first release.
This one still has watermarks in the corners of a few input images, which are easy to prompt out but should all be context-aware-filled by next release.
Trained on ChilloutMix for 400 impressions per image (25 repeats X 4 epochs X 4 gradient accumulate steps), batch 6, Lion optimizer, 0.00005 unet LR, 0.00002 tenc LR. 207 manually-tagged images in 6 categories. This was an intermediate save from the ~26th footjob model training test, and the first release.
This one still has watermarks in the corners of a few input images, which are easy to prompt out but should all be context-aware-filled by next release.