This is an experimental merged model that incorporates an Asian Lora I retrained.

Tips:The checkpoint is compatible with other Flux.1 Dev Loras, but you might need to tweak the Lora weights a bit.

After multiple adjustments, the model retains a good variety of Asian facial features with some aesthetic tweaks. Out of around 1000 test runs, it performed pretty well overall, though a few issues were noticed.

Despite my efforts to minimize the impact on other parts of the model, the overall brightness slightly increased, which may lead to some differences in composition compared to the original model. The overall performance is still solid, but unfortunately, there's been some minor degradation in limb accuracy. Hopefully, I'll find a better solution in the next version.

Version options:

The FP16 version is more precise than the FP8 version, so it offers better image quality and fewer issues with limbs. The chance of limb distortion is significantly lower.

The FP16 version (~22GB) is about twice the size of the FP8 version (~11GB).

The FP16 version takes up more VRAM. If you’re running low on VRAM, the FP16 generation time will be much longer than FP8, but if you have enough VRAM (>=24GB), the speed difference is minimal.

Note: This is just the model(unet) itself, it doesn’t include vae, clip, or t5xxl.