Note: I don't plan on releasing another version of this model; there are already so many good photorealistic models out there, adding more just creates clutter! But I like having my own and I am putting this out there so other people can faithfully replicate any images that I create with it.
This pet project was my custom, mostly-photorealistic model, blended and fine-tuned on hand-captioned images of people, primarily from professional photo shoots. My goal was to create a model that generates beautiful images of people by default without needing excessive prompts, so that prompts can focus on adding qualities instead of removing defects.
When you're going for photographic images, you may wish to place "illustration" in the negative prompts. The following negative prompts are usually good: "3D digital illustration, blurry, low quality, grainy, pixelated" But remember to take "illustration" out of the negative prompt if you're going for art/anime/comic style results.
Description
FAQ
Comments (5)
How is it with intricate poses? legs up/spread eagle/yoga poses. Is this a model training problem or just a stable diffusion problem?
I think it will depend on the pose. Spread legs works fine (in fact, I think it's one of this model's preferred default poses), and you can manipulate poses through Control Net. I'll run some tests and see if I can pinpoint some specific limitations. Do you have any prompts in mind that you're curious about?
@Invisidude prompting for complex exotic porn poses never works in stable diffusion, so I just use img2img or controlnet. Ive yet to find a model that gets "spread eagle" right without mutated legs. I think its just a stable diffussion issue.
@fox23vang226 yeah, I think at this point our best bet for specific poses is either Control Net or dedicated pose LoRAs. I'll keep trying to make the base model better, though!
@fox23vang226 its SD issue, simply SD doesn't have any concept how human body should actually look and how it bends that why we have so many issues with missing limb, too many limbs (it dosnt do math too). Its why controlnet is so useful as it guides SD to what we want.



