Overall, use Flux (h/t Fireship), train on fal.ai, then generate via API/web app to bypass safety filter.
Cost: ~$5. No GPU needed!
- Collect 10-20 nude images and zip them into a folder
- Use “Train a LoRA” on civitai.com, steps 1 & 2, to generate captions via Auto Label
- Clean up the captions (see this), then download the captioned images as a new zip.
- Upload the zipped dataset to https://fal.ai/models/fal-ai/flux-lora-fast-training.
- Higher iter multiplier (eg 3) increases faithfulness to character
- Once completed, copy the
diffusers_lora_file.url - Clone
Embed GitHub and modify to accept LoRA url - (TODO: upload my fork?)
Further reading:
- https://civitai.com/articles/7097/flux-complete-lora-settings-and-dataset-guide-post-mortem-of-two-weeks-of-learning
- Try runware.ai? Not clear if it supports Flux LoRAs