Thx. and nice work
But how can I use the models? Just copy them into the models folder? ... or what?
How get you so nice repainting? My img2img is always fucked up ^^
Put the models in models/stable-diffusion, then restart stable-diffusion-webui and you can select them in the upper left corner.
To get a good image I found you need a good negative prompt, I use this:
Code:
black and white,border, frame, lowres, jpeg artifacts, low quality, normal quality, worst quality, poorly drawn, error. abstract, asymmetrical, blurry, cropped, disconnected, duplicate, extra, missing, signature, text, username, watermark. amputee, bad anatomy, deformed, disfigured, disproportionate, eye bags, fused, malformed, morbid, mutated, mutation, mutilated, hands, fingers,bad hand,extra limbs,missing limb,extra leg,missing leg,missing hand,extra hand, barefoot, feet, toes, bad feet
but you can also google around for pages where you can see ai art with prompts just to get you started, also on that
You must be registered to see the links
site, the example pictures in the models sometimes have a little i icon in the lower right corner, you can click that to see the prompt and parameters that were used to create the image. This might be better then random prompts you find because it seems the models can require quite specific prompts/parameters to work well.
For the prompt itself you can also click "Interrogate CLIP" in the img2img tab. This will use image recognition to try and figure out what your image contains and use that as a starting point for your prompt. I found that very often this is enough for img2img as long as I have a negative prompt as well.
It is very important you actually use the width and height sliders to fit your image well. If you don't then the model will not work correctly. I found I often get like multiple mouths and eyes when I forget to use the width and height sliders after changing the image.
Finally the two most important sliders to control the result are "CFG Scale" and "Denoising Strength" on the left side of the img2img tab at the bottom.
The "CFG Scale" controls how much your prompt will influence the picture, I found low values there to be very good, maybe because I often just use a generic prompt and don't want to change the input image too much. Higher values in "CFG Scale" tend to make the image changes more drastic in a way, though only color and contrast, not so much structure. I am not really sure I understand this yet, I tend to think it is like with a prompt that says "women with red hair", a high value means "I want it very woman and with very red hair please", whereas a low value means "red-ish is fine, but light-brown hair is great too". Kinda like that.
Otoh "Denoising Strength" controls how close the resulting image stays to your input, or how much "dreaming" you will allow the AI to do. Again I prefer low-ish values there, but high values are more fun! A low value will just apply a styling to your image, keeping the structure intact. A higher value will create a more and more differently structured picture, which sometimes can result in really cool stuff but also in really broken stuff.
Code:
a woman in a skirt and shirt and red hair standing in front of a blackboard with a table and a chair
Negative prompt: black and white,border, frame, lowres, jpeg artifacts, low quality, normal quality, worst quality, poorly drawn, error. abstract, asymmetrical, blurry, cropped, disconnected, duplicate, extra, missing, signature, text, username, watermark. amputee, bad anatomy, deformed, disfigured, disproportionate, eye bags, fused, malformed, morbid, mutated, mutation, mutilated, hands, fingers,bad hand,extra limbs,missing limb,extra leg,missing leg,missing hand,extra hand, barefoot, feet, toes, bad feet
Steps: 30, Sampler: Euler a, CFG scale: 6, Seed: 3363671156, Size: 704x1088, Model hash: e2bc6c56, Denoising strength: 0.39, Mask blur: 4
EDIT:
Code:
a woman in a skirt and shirt and red hair standing in front of a blackboard with a table and a chair
Negative prompt: black and white,border, frame, lowres, jpeg artifacts, low quality, normal quality, worst quality, poorly drawn, error. abstract, asymmetrical, blurry, cropped, disconnected, duplicate, extra, missing, signature, text, username, watermark. amputee, bad anatomy, deformed, disfigured, disproportionate, eye bags, fused, malformed, morbid, mutated, mutation, mutilated, hands, fingers,bad hand,extra limbs,missing limb,extra leg,missing leg,missing hand,extra hand, barefoot, feet, toes, bad feet
Steps: 30, Sampler: Euler a, CFG scale: 6, Seed: 2235074263, Size: 704x1088, Model hash: e2bc6c56, Denoising strength: 0.8, Mask blur: 4
The second one is from the input image, but you can see how a higher "Denoising strength" makes it so the model starts taking things into its own hands and just goes wild with your input. This also took several tries to get one where the result looks like a normal person (two arms, two legs).
EDIT2:
And let me link the models again I use, cause they are so important:
You must be registered to see the links
(already linked this one and seems overall the best I could find so far to just make sexy/cute woman)
You must be registered to see the links
(similar more anime-ish result like the pfg one, looks more drawn then rendered imho)
This is Berrymix with the input and params from my first EDIT (Seed is 1538804157 though). This is so fun actually, I can't even stop adding more stuff to this post.
