NOTE :
Base model : Z Image Turbo & Flux1.
Showcase images were created with ComfyUI and are in PNG, include metadata & workflow.
Description
Model trained on civitai with a base of 77 images
Recommended resources : Fluxmania IV
Recommended settings : dpmpp_2m or heunpp2 / sgm_uniform / 25 steps / guidance : 2.5 to 4.0
Weighting : around 0.5
FAQ
Comments (2)
I get very poor results, I think roughly 1 in 10 is somewhat usable. I tried both F1D and Fluxmania IV as checkpoints and various to no LoRAs. Pity, because I really enjoy your LoRAs
Thanks for your feedback. I also experienced some anatomical aberrations sometimes, but much less frequently, the frame rate was largely acceptable. However, I noticed that certain postures and positions had a very high aberration rate, such as a lying position, or postures where the hands and legs are intertwined. I think these aberrations originate from the FLUX model itself.
Regarding the aesthetic quality of the lighting, composition, etc., the results seemed very correct to me.
Unfortunately, you can never be sure to get a model as efficient as you want. I have about 20% of trained models that I failed to correct and were unshareable.
For this one, I thought it could be shared, even with a lower success rate than the rest of the models