<p>Character LoRA trained on Big Love Photo 5. For fun, I created a one-click* character LoRA generator app. It uses LM Studio to create the character description and background, sends that to ComfyUI to generate the initial image with BigLove Photo 5, then uses Comfyui again wtih Klein 9B FP8 KV to create the dataset images from the initial images, adds the dataset prompts, then passes all of that off to kohya_ss to train the LoRA. This is the first one I created after tuning everything and fixing bugs. Seemed to turn out okay. Could use any other base SDXL model easily enough. The dataset is SFW, but does well with NSFW when used with BigLove. Whole process takes about an hour on my Ryzen AI Max+ 395. Probably wouldn't work too well as-is on anything with less than 96GB of VRAM.<br /><br />* there's a lot of initial setup and a few regenerating bad dataset images. Could just let it run on its own after the setup, but you may have bad results if the dataset images have bad limbs.</p>

Files with SHA256: 108ac8a9a6f1726051ef6b23bb702ea5142ef226c9ca4d68686e04dbb84af038