CivArchive
    mini-pixel-art - S-P1
    Preview 149438066

    mini-pixel-art — tiny 16×16 pixel-art sprite generator

    Not a Stable Diffusion / SDXL / Flux model. It will not load in ComfyUI, A1111 or Forge. It is a standalone 1.9M–8.8M-parameter diffusion model that draws native 16×16 sprites with transparency, directly in pixel space (no VAE, no CLIP).

    ▶ Try it in your browser (WebGPU, nothing uploaded): https://anticpan.github.io/mini-pixel-art/

    Code: https://github.com/AnticPan/mini-pixel-art

    All weights + ONNX: https://huggingface.co/AnticPan/mini-pixel-art

    What it does

    - Text → 16×16 RGBA sprite. Output colours come from a fixed 4,096-colour palette, and the background is real transparency.

    - Fast enough for games and tools: ~130 ms per image in a browser (M-P1, WebGPU), ~250 ms on 4 CPU threads, ~870 images/s on one RTX 5090.

    - Files are tiny: 4.5 MB (int8) / 17 MB (fp32) for the recommended model.

    Versions

    Version | Params | Notes

    M-P1 | 4.35M | Recommended

    L-P1 | 8.75M | Best quality, ~1.8× slower

    S-P1 | 1.9M | Smallest, weaker at hitting the subject

    M-P2 | 4.36M | ~2.5× faster than M-P1, a bit less detail

    How to use

    python mini_pixel_art.py --model m-p1 --n 8 "red potion bottle" "side-view knight with sword" --out sheet.png

    (PyTorch, or mini_pixel_art_onnx.py with onnxruntime only.) Defaults: 16 steps, guidance 1.5.

    Prompting

    - Short noun phrases like the training captions: golden key, wooden treasure chest, blue mage hat, green slime.

    - Optional view word at the start: front-view, three-quarter-view, top-down, side-view.

    - Closed vocabulary (~11k words from the training captions); unknown words are ignored. No negative prompt, no LoRA, no img2img.

    ## Limitations

    - Softer than hand-made pixel art; fine detail is limited.

    - Flat colour areas often use several near-identical shades (about 50 colours per sprite vs ~22 in real sprites).

    - Only knows subjects from its training data; it does not generalise to new concepts.

    Training

    Trained from scratch on [Angais/pixel-art-atlas-1.4m](https://huggingface.co/datasets/Angais/pixel-art-atlas-1.4m) (1.4M synthetic 16×16 sprites, CC0). According to the dataset card, those sprites were generated with Krea 2 Turbo + a pixel-art LoRA. Flow matching, 300k steps, batch 512, one RTX 5090.

    License

    MIT for code and weights. The weights contain no part of any other model. See the NOTICE file for data provenance (including the dataset card's note on Krea 2 terms). Not affiliated with Krea or the dataset author.

    Description

    Checkpoint
    Other

    Details

    Downloads
    5
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/18/2026
    Updated
    11/20/2026
    Deleted
    -

    Files

    miniPixelArt_sP1.safetensors

    Mirrors

    HuggingFace (1 mirrors)
    CivitAI (1 mirrors)