Skip to content
Images

Pixie P32

a compact AI model for 32×32 pixel art sprites

2 min readupdated

Created by EmilioSchwaiger, this model generates native 32×32 pixel art sprites with transparent backgrounds from short English text prompts. Unlike standard diffusion models that output high-resolution images requiring downscaling, Pixie P32 is a 66.4M-parameter autoregressive transformer that predicts palette color indices directly across 2×2 patches. For text conditioning, it uses the first 11 layers of the external openai/clip-vit-large-patch14 encoder. It is built for indie game developers and researchers who need rapid prototypes of game icons, items, or creatures without running heavy generators.

Key features

  • Native pixel art: generates directly on a 32×32 grid using a 4,097-color palette (index 0 maps to RGBA transparency), avoiding blur and compression artifacts.
  • Lightweight architecture: 16 layers, 512 hidden width, 8 attention heads, and partial 2D-RoPE (positional encoding for 2D grids), totaling roughly 66.4M parameters.
  • Style and palette slots: each 2×2 patch is encoded using three palette slots and one style slot.
  • Permissive license: model weights are released under Apache-2.0, with training and inference source code available on GitHub.

What you can do

  • 2D game prototyping: quickly batch-generate inventory icons—potions, swords, helmets, coins, or level props.
  • Character and enemy drafting: create base sprites for creatures (such as slimes or monsters) ready for manual cleanup by an artist.
  • Research and fine-tuning: use the pipeline code from the QuarkOS/pixie repository to experiment with autoregressive indexed-graphics generation.

Example

hf download EmilioSchwaiger/pixie-32 --local-dir pixie-32
hf download openai/clip-vit-large-patch14 model.safetensors tokenizer.json --local-dir clip-vit-large-patch14
python -m train.sample_pixie --model-dir pixie-32 \
  --clip clip-vit-large-patch14/model.safetensors \
  --tokenizer clip-vit-large-patch14/tokenizer.json \
  --prompt "red potion" --prompt "iron helmet" --n 4 --out sprites.png
Read moreDetails
Metrics: IID validation NLL reached 0.5237 at step 17,746 (roughly 1.25M images seen).
Training: trained on a dataset of ~130k 32×32 sprites generated via Krea 2 Turbo with a pixel art LoRA.
Requirements: requires a separate download of the CLIP ViT-L/14 text encoder (not bundled with the weights).
Limitations: experimental research model with inconsistent output quality; magenta fringe pixels from the training chroma key occasionally appear along sprite edges.
Discuss Telegram 1 36

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 0.2 GB · ready files1.8 GB
  • BF16weights 0.1 GB1.6 GB
  • FP8 / INT8recommendedweights 0.1 GB1.6 GB
  • 4-bitweights 0.0 GB1.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · F32 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # EmilioSchwaiger/pixie-32
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download EmilioSchwaiger/pixie-32 --local-dir ~/models/pixie-32
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to pixie-32
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • SageAttention 3v2.2.0from October 28, 2025

    even faster on RTX 50 / Blackwell

    FP4 attention for Blackwell. Faster than SageAttention 2, but the authors suggest v2 when accuracy matters.

    How to enable: Build from the sageattention3_blackwell folder of the repo; in ComfyUI via KJNodes.

    Earlier generations (2)
    • SageAttention 2++ — Quantised (INT8/FP8) attention with no visible quality loss. The main speed-up for image and video generation.
    • SageAttention 1 — First version: INT8 for Q·K. Still works, but 2++ is faster and more accurate.
  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (3)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
    • xFormers — The older way to speed up attention, works almost everywhere. Sage or Flash are faster where available.
  • cache-dit (DBCache)v1.5.6from November 20, 2026new release

    1.5–3× at nearly the same quality

    Reuses block outputs between diffusion steps. Supports most DiT models.

    How to enable: pip install cache-dit; one line cache_dit.enable_cache(pipe) for diffusers.

    Earlier generations (2)
    • MagCache — Skips steps by magnitude of change. More accurate than TeaCache, no calibration.
    • TeaCache — The first popular diffusion step cache. Works, newer methods are more accurate.
  • Nunchaku (SVDQuant 4-bit)v1.2.1from January 25, 2026

    ≈3× faster and 3.5× less memory

    4-bit versions of popular image models with almost no quality loss (FP4 on RTX 50).

    How to enable: Ready weights for supported models; in ComfyUI use ComfyUI-nunchaku.

  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of pixie-32 (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

EmilioSchwaiger/pixie-32
Author
EmilioSchwaiger
Category
Images
Released
Posted

On NeyroLearn

Hugging Face

361

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments