(Skormino) (2025)
Image description: A pixelated composition featuring a vibrant, pink flower in the foreground, standing tall with a green stem and leaves. The flower is surrounded by a field of pink flowers that stretches to a curved horizon at the bottom half of the image. In the background and to the right, a bright blue and green planet is partially visible, with a larger dark planet looming behind it. A bright star peeks out from behind the larger planet in the top right corner, illuminating the dark planet's atmosphere.
Full Generation Parameters:
masterprice, , dithered background, flower, sky, petals, no_humans, black_background, red_flower, star_\(sky\), scenery, pink_flower, starry_sky, space, field, planet, flower_field, earth_\(planet\), still_life
Negative prompt: (monochrome, sketch:1.2), watermark, muscle, abs, ugly, deformed, noisy, blurry, low contrast, drab, boring,
Steps: 42, CFG scale: 4, Sampler: Euler, Seed: 16425, Size: 1024x1536, Model: noobai\yiffymix_v62Noobxl.ckpt, Batch size: 2, Model hash: 2b347f759a, Batch index: 0, Lora_0 Model hash: 7c9c745b5d, Lora_0 Model name: pixpix_4_r1.safetensors, Lora_0 Strength clip: 1, Lora_0 Strength model: 1, Clip skip: 2
I wonder if a model could actually do pixel art the "classic" way. It wouldn't take much processing to identify pixel grids in existing pixel art and convert them to native resolution for training.
Yes: they train the model on 8 bit art with 8x8 pixels and then ask it to generate with a scale of 0.125 (1/8) and scale up with no interpolation.
Check out the model page for an explanation on how it works. It seems to mostly get simpler scenes right. The gallery below the comments has some nice images too.