Regarding my next realistic model, I’ve long wanted to try creating a Flux-style Illustrious checkpoint. So, I used around 100,000 (or 86,000?) images generated by Flux to develop an upgraded LoRA for my illust-2.8Dmix. Currently, the training has reached Epoch 3, so I merged them for testing.
The results are quite unstable: it’s not ideal for low steps(within 25 steps) image generation, as the success rate of generated images has dropped significantly and requires AD repair. That said, the quality of the generated images is excellent, The depth of field, color tone, and picture style are all very similar to flux.
Now I’m hesitant—should I release it as V4, or as a separate test model instead? or continue to improve it?
