About

I proudly introduce you to Famira, my first Flux Lora that is based on a set of merged faces, which turns out to be a very average Asian woman. Flux.1 Dev is really astonishing, but it still contains render flaws. I am not using the best Workflow setup, but hey... maybe you guys may do better.

Training

First time I rendered a Lora was on Fal.ai, but due to limited parameters, the results was really nothing like the Civitai Lora trainer. At Fal.ai my model was probably also overtrained. It seems like for Lora faces it requires about 2000-ish steps.

Famira is trained with 45 (square 1024) with (5 repeats, 4 batches), which means... 45 x 5 repeats / 4 batches x 36 epochs = 2025 steps.

About rendering with Flux.1 Dev

Well, what can I say, they really hate cigarets. I tried to render Famira smoking, but still the hands are an issue. Probably requires different Lora. For the rest, I really hate the fact that it is missing negative prompts. Maybe I am doing something wrong, in order to fix that, I sometimes have to use false positive prompt.

If you like my work, don't hesitate to leave a like or a review. Please share your renders so I gain knowledge what other potentials this model can be used for.

Last but not least, I drink a lot of Ko-Fi :)