- cross-posted to:
- asklemmy@lemmy.ml
- cross-posted to:
- asklemmy@lemmy.ml
Hello,
I’m trying to run the pixel art generator (https://perchance.org/ai-pixel-art-generator) locally on my machine so that I can run it programmatically from Python. From what I’ve gathered (mainly from this post: https://lemmy.world/post/5926365), the model behind the generator is SD 1.5, however, I’ve tried running it locally (downloaded from https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main, tried both emaonly and pruned) and I can’t seem to be able to replicate the results. It would be helpful to know the exact configuration used to prompt SD 1.5, or provide some more help on how to set up some API, or even point to some Github with the code that I haven’t been able to find so that might not exist. I’ve tried to read all the documentation I could find but was not able to use any of the provided resources (like this one: https://perchance.org/diy-perchance-api, or this one: https://perchance.org/text-to-image-plugin). If anyone could help with any of the above, I would be eternally grateful <3. I will also list everything I’ve pieced together so far on how the generator works in case someone else might find it useful. The images generated with this configuration are similar in style to what pops out of the generator but they are just fundamentally different in quality. The ones from the generator are much better at depicting the prompt.
What I’ve gathered so far:
- Model used: SD 1.5
- Width x Height: 512x512
- Sampling method (not a clue what this should be): DPM++ SDE
- Prompt: <prompt>, best pixel art, neo-geo graphical style, retro nostalgic masterpiece, 128px, 16-bit pixel art , 2D pixel art style, adventure game pixel art, inspired by the art style of hyper light drifter, masterful dithering, superb composition, beautiful palette, exquisite pixel detail
- Negative Prompt: glitched, deep fried, jpeg artifacts, out of focus, gradient, soft focus, low quality, poorly drawn, blur, grainy fabric texture, text, bad art, boring colors, blurry platformer screenshot
What models do you use?
I’ve been testing most stable diffusion models, 1.4, 1.5, 2.1, nano, I just tested XL. None seem to really match the images I get from the website
Have you tried any of the pixel fine-tunes on Civitai?
This one looks pretty good.
I enjoy this 1.5 Lora https://civitai.com/models/165876/2d-pixel-toolkit-2d it’s pretty neat!