This LoRA was trained on approximately 15 images of Gal Gadot, initially upscaled using SUPIR Upscaler. An initial upscale was also performed using another previously trained LoRA with the images outputted by the SupIR upscaler, using a denoise strength of 0.3 and applying "ultimate SD upscaler" on the Flux Dev diffusion model.

For optimal use, I recommend the following settings:

• Sampler: Heun

• Scheduler: Beta

• Steps range: between 20 and 50.

• Strength guidance: 3.5

This LoRA exhibits high text adherence and versatility in facial expressions. I recommend using the exact same input for the CLIP L text encoder as for the t5xxl text encoder; I've conducted dozens of tests, and it performs better with longer text prompts.

The LoRA's style is primarily realistic, with strong compositional capabilities. I have trained another LoRA (with 15 images) that I consider superior to this one in terms of aesthetic and visual quality (in all tests using computer vision models, the other LoRA, which I'm not releasing, consistently scored higher in aesthetics and quality). However, I won't be releasing it today, perhaps later (it's inferior in facial representation but superior in representing Gal Gadot's physique).

This LoRA was trained with a lora rank of 64 on an RTX 4090, so it's not exactly accessible to everyone. That said, the realism it delivers surpasses many other Gal Gadot LoRAs.