The details are embedded in the png. In A1111, you can switch to the PNG Info tab and load the image to see and reuse them.
Here they are for everyone:
wallpaper, detailed background, Intricate, dramatic, photorealistic, majestic oil painting, beautiful artwork, sunset, full body shot, german woman, swedish woman, blonde, Ed Blinkey, Atey Ghailan, Studio Ghibli, Jeremy Mann, Greg Manchess, Antonio Moro, greg rutkowski Negative prompt: sketches, lowres, bad anatomy, bad hands, missing fingers, poorly drawn, (out of frame), (cropped), ((worst quality)), ((low quality)), jpeg artifacts, text, error, signature, watermark, username, artist name, blurry, out of focus, ((monochrome)), ((grayscale)) Steps: 32, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1511898655, Size: 512x384, Model hash: 9aba26abdf, Model: deliberate_v2, Denoising strength: 0.5, Hires upscale: 2, Hires upscaler: Latent
The names are artists, to guide the style. The rest should be fairly self explanatory. Parentheses amplify the parameter.
Negative prompts are important to have, to filter out unwanted things. Note that they act as a filter on source material; they don't erase or avoid things in the output (many people misuse them, not understanding the distinction).
You might notice there's surprisingly little to guide the actual content: "sunset, full body shot, german woman, swedish woman, blonde"
Every other detail is dreamed up by the AI, indirectly guided by the style tags, the model, the sampler, and the random seed. Sometimes it pays off to generate multiple images with the same tags, but different random seeds. The sampler used also has a massive impact on the result.
Yeah, maintaining character is a tough problem for these models. They don't have any kind of facial recognition or concept of identity built in. They're just dreaming up images based on their training data and the prompt provided, plus some random noise.
I'm not aware of any good solutions for it yet. Your best bet might be using a celebrity (or combination of celebrities) to get some consistency. Or be super specific about the facial features in the prompt and negative prompt.
Reading up on it a bit, it looks like the latest and greatest way to do what you want is using "embeddings" and "textual-inversion". That involves training a small model (or finding a suitable one) which effectively gives the main model an understanding of a custom concept (which can be a person) that you can then reference in the prompt.
[ + ] Sleazy
[ - ] Sleazy 0 points 2 yearsApr 23, 2023 02:20:06 ago (+0/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 0 points 2 yearsApr 23, 2023 02:28:55 ago (+0/-0)
[ + ] Sleazy
[ - ] Sleazy 0 points 2 yearsApr 23, 2023 04:07:41 ago (+0/-0)
[ + ] Deleted
[ - ] deleted 1 point 2 yearsApr 23, 2023 00:10:24 ago (+1/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2 yearsApr 23, 2023 00:32:34 ago (+1/-0)*
Here they are for everyone:
Negative prompt: sketches, lowres, bad anatomy, bad hands, missing fingers, poorly drawn, (out of frame), (cropped), ((worst quality)), ((low quality)), jpeg artifacts, text, error, signature, watermark, username, artist name, blurry, out of focus, ((monochrome)), ((grayscale))
Steps: 32, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1511898655, Size: 512x384, Model hash: 9aba26abdf, Model: deliberate_v2, Denoising strength: 0.5, Hires upscale: 2, Hires upscaler: Latent
The names are artists, to guide the style. The rest should be fairly self explanatory. Parentheses amplify the parameter.
Negative prompts are important to have, to filter out unwanted things. Note that they act as a filter on source material; they don't erase or avoid things in the output (many people misuse them, not understanding the distinction).
You might notice there's surprisingly little to guide the actual content: "sunset, full body shot, german woman, swedish woman, blonde"
Every other detail is dreamed up by the AI, indirectly guided by the style tags, the model, the sampler, and the random seed. Sometimes it pays off to generate multiple images with the same tags, but different random seeds. The sampler used also has a massive impact on the result.
[ + ] Deleted
[ - ] deleted 0 points 2 yearsApr 23, 2023 01:37:40 ago (+0/-0)
[ + ] Deleted
[ - ] deleted 0 points 2 yearsApr 23, 2023 03:09:22 ago (+0/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2 yearsApr 23, 2023 03:38:11 ago (+1/-0)
I'm not aware of any good solutions for it yet. Your best bet might be using a celebrity (or combination of celebrities) to get some consistency. Or be super specific about the facial features in the prompt and negative prompt.
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2 yearsApr 23, 2023 04:03:45 ago (+1/-0)
https://rentry.org/embeddings
https://rentry.org/textard
"Dreambooth" is also something that gets mentioned, but I haven't looked into that at all yet.
[ + ] observation1
[ - ] observation1 1 point 2 yearsApr 23, 2023 00:28:43 ago (+1/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons [op] 1 point 2 yearsApr 23, 2023 00:40:20 ago (+1/-0)