@john Something like this? The trickiest bit was getting it so it doesn't look photoshopped. The inpainting was denoised at 0.9. I don't know if the Americanisations made any difference, I suspect not. There is cleaning up to do around the tail and legs, but that is to be expected with #stablediffusion. The initial model was SD's 768-v-ema and the inpainting used the Realistic Vision 5.1 inpainting model, masking the eyes and muzzle.
@BlackPhi Yeah, infilling is possible, but I guess what I'm interested in is not so much how to get a certain thing done (I can paint/photomanipulate/etc. myself), but what is going on with AI imaging in general .
I'm finding the coaxing aspect interesting. It's just not as simple as the model doesn't have reference images, because DALL-e makes some things hard, but with some cajoling will actually do a pretty good job.
@akira28 Yes, if you can load the rest of the thread (unreliable on Mastodon, I know), you'll see that ChatGPT gives a more elaborate prompt to Dall-E, based on your prompt. If you use Bings version you need a fairly elaborate prompt to get it to work.
@john I'm insufficiently motivated to spend my own money buying DallE credits in service of this question, but I wonder how it would handle the prompt "A sphinx with the body of a fox"
@john Furry porn indeed, lol. I was hoping it would know a sphinx has the head of a human and the body of a lion,and thus understand it was supposed to keep the human head while replacing the lion parts with fox. Instead it went full on Egypt while keeping the fox face. A centaur with the body of a fox isn't quite what you want, since I assume you don't want the human torso. Hmm. Interesting challenge.
@john another failure "a photograph of a man caught outside in a realistic fox costume, with his shocked face looking directly at the camera, in the style of nature photography"
@zillophane And God prompted, “let there be a fox, with the face of a fox, and yet of a man, so that no man knoweth why it unsettles him so" and so it was.
A realistic orange vulpine lammasu, with the body of a fox. The head belongs to Albert Einstein. The fox is orange. The head is smiling. The body is visible. The tail is bushy. The head is a scientist with his tongue out. The face is pink. The head is fully human.
And negative prompt: fox-headed
The vague idea was to get a well-known human head in there to force the AI into using it.
@TEG Dall-e mini, apparently, which is an attempt to match Dall-e with an open source model. It's seems to pay more attention to the prompt than Stable Diffusion, which gave me this:
@john@AggroBoy ChatGPT does do some stuff behind the scenes, including writing its own prompt. You can see part of that in the filename when you download the image. The prompt it wrote for the first image I posted was
"A surreal and imaginative depiction of a fox with a human head, blending the natural orange and white fur of the fox with the distinct features of a h..." (the filename cuts off there).
@john@pbloem@AggroBoy At least ChatGPT, because it is trained for that purpose. But it can only talk to dall-e, I guess other image generation ais don’t like the generated prompts very much.
@john I spent about half an hour trying last night and just couldn't get it to do it. In a great many attempts, I got one example out of DallE4 (that I didn't save) where it had sortof superimposed a textureless grey human face on top of a fox's head, but it didn't look like it was actually part of the head. The rest were either normal foxes, or humans wirh fox heads. Weirdly, it really liked generating humans with HUGE foxes heads.
Add comment