My best guess for your training set would be Set753.
I dare say nobody would object to nsfw photos, but I think they tend to look more obviously fake, as the models don't seem to be as effective for body shapes.
I tried to train a model for Silvia on whole body photos and it took days on a RTX 3050 laptop. My first effort to generate an image gave her three legs, so I promptly gave up!
I used only close-up photos of the face to train the model. I have not studied how to copy the body and shape, all the generation is due to the basic model (but it can be adjusted with prompts). The generated photos also have a plastic structure. But I think if you study deeper, you can achieve better results.
-- Edited by sam917 on Wednesday 17th of June 2026 03:16:54 PM
-- Edited by sam917 on Wednesday 17th of June 2026 04:56:46 PM
When the face is turned to the side, the quality is worse than when it is looking straight. I have uploaded all the good images. Are there any images that you like?
@sam917: After seeing what you've achieved, I've been playing around with Koyha_ss and Stable Diffusion too
I've gone for a younger version of Silvia.
I got nowhere with SDXL Lora models - huge training times with rubbish results. Going back to a basics with a standard SD 1.5 Lora seems to work a bit better. I've still not managed to achieve your level photorealistic quality, but it's showing some some potential. This is the best I've managed so far:
For the record, this Lora training was done in Koyha_ss using pre-trained model "runwayml/stable-diffusion-v1-5" and generated in Stable Diffusion using "v1-5-pruned-emaonly.safetensors" checkpoint.
Still some way to go, so I'll continue trying to refine the model.
As an aside, the Koyha_ss SDXL training threw up some interesting samples during the 8 hours or so it took to generate an ultimately disappointing result. Here's one of what a 50 year old Silvie Tomalová might look like (but probably not):