This LORA is made for messy ??????? photos with full facial and body coverage. The LORA aims for ??? coverage over all else.

FOR flux-dev-fp8 users: this LORA does not function well on the quantized version of flux. Instead, check out Cumzilla 8-bit Flux Edition which is an fp8 quantization of a fine tuned version of Flux based on the Cumzilla dataset.

This is v4 of Cumziilla. I was never a huge fan of v3 (only 50 epochs of training) and went for 100 this time with more concise joy captions that were human-reviewed and edited. I also went from rank=4/alpha=16 to rank=8/alpha=32 to see if it makes a difference. v2 is 150 epochs if you want to try something stronger.

The LORA needs a large amount of prompting help to render ??? properly. See example images for help with prompting - e.g. if you don't tell it ??? is white, you're going to get lots of random colors. See example images for prompting advice.

Changes

• Using more concise and human-edited/reviewed Joy captions (highly descriptive) for training - don't notice much of a difference

• Pruned some blurry images for higher qualiity

• Increased steps to 100 epochs vs 50 in v3. v2 was trained at 150, so there are some different strengths you can try out in your experiments.

• Trained on the latest/greatest LORA config.

NOTE: this LORA performs poorly against the flux1-dev-fp8 quantization of Flux. Don't be surprised if results are off. It was trained on and performs the best on flux1-dev. Not sure how to help this in terms of training. One user suggested using Heun with 8 steps for the fp8 quantized Flux.