Workflow for consistent faces:

I combined 2 loras. (e.g. <lora1:x><lora2:y><lora3:z> x+y+z=1 experiment till consistency and aesthetic) Others have pointed out faceswapper is quicker. I havent tried.

Generated a batch without loras and selected a image I liked (prompt ethnicity, hair style and length for consistemcy). Recycled seed and ADetailered face with loras and controlnet.

Workflow for text on card: SD will generate blank paper. Physically write what you wamt on a paper, more textured, the better. Layer as hard light in photoshop and clean up.

Workflow for text on skin: Write using brush in photoshop. Warp to match body contour. Run img2img inpaint with paint on skin in prompt, high denoising, with Controlnet Canny. May have to crop to 512x512 and enlarge text area before img2img for good results. Again layer image in photoshop.


Why am I adding LoRas up to 1?

Easier to calculate ratios and tinker using x/y/z plot. Too much cumulative LORA weight reduces imaginative ability of checkpoint.