Moss Image for pixel art
retained text
I enjoy starting godot game projects that I never finish. While I very much enjoy carefully plotting pixels in Aseprite I have also looked at AI to aid me in that process. I've been trying pixel art with MOSS-Image-SFT and noticed that it produces a fairly decent pixel grid straight out of generation, before any snapping or cleanup. It also seems to have a good general understanding of pixel art. That got me curious about what a small LoRA could do with it. The experiment is mostly about making the pixel scale more predictable. I trained on 30 pixel-snapped 64×64 sprites enlarged to 512×512, targeting an **8× scale** where each logical pixel occupies an 8×8 block in the generated image. That gives these intended working sizes: |Generate at|Intended pixel grid| |:-|:-| |512×512|64×64| |768×768|96×96| |1024×1024|128×128| It's still a small experiment. The grid can drift, edges can bleed, and some details won't survive downscaling. The 8× scale is a training target, not a guarantee that every output will land perfectly on it, but I'm still surprised on how close it is. Here is a diff between output and grid snapped and color quantized image: [https://www.diffchecker.com/image-compare/ateIcHYx/](https://www.diffchecker.com/image-compare/ateIcHYx/) That downscale doesn't fix uneven cells. For additional cleanup I use [Sprite Fusion Pixel Snapper](https://www.spritefusion.com/pixel-snapper). It can help recover a consistent grid. It also exist as a [Comfy node](https://registry.comfy.org/publishers/x0x0b/nodes/comfyui-spritefusion-pixel-snapper). Model can be downloaded from [CivitAI](https://civitai.com/models/3024672/moss-pixel-art) or Hugging face if you want to play around with it. Workflow and training details can be found in the [Hugging Face](https://huggingface.co/finger-gun/moss-pixelart-v.0.3) repo.
document fetched 2026-10-29 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1xhynoq/moss_image_for_pixel_art