Kind of an academic experiment but maybe someone will find it useful. The workflow I used utilizes a chroma – Z-image handoff directly through latents without decoding. This has a few benefits. First, Chroma knows a whole lot of NSFW, so letting it define the structure of the image opens up a lot of possibilities. Z-image as a refiner still shreds penises, but I imagine that will change very soon now that people are making LoRas. In the meantime, I also have a workflow that uses segmentation to detect penises in the chroma image and then uses the Chroma model to fix the end result. Works so-so, misses detection a lot.
Workflows are here: Chroma-Z-Image + Controlnet workflow | Civitai
Models used are listed in the workflows.
Article Categories:
unstable_diffusion