Pixie P32 (pixie-32)

A 66M-parameter autoregressive model that paints native 32ร—32 palette-indexed pixel-art sprites from a short text prompt. Code, training pipeline and sampler: github.com/QuarkOS/pixie. Sibling model: EmilioSchwaiger/pixie-16.

Files: model.safetensors (weights), config.json (architecture), palette.npy (4097ร—3 uint8 palette; index 0 is transparent).

Samples: red potion, iron helmet, green slime (4 each, seed 0, CFG 2.0)

Model

Field Value
Name Pixie P32 (pixie-32)
Version Best checkpoint of the first full P32 run: step 17,746, about 1.25M training images seen
Architecture P32: autoregressive transformer over palette indices. Width 512, 16 layers, 8 heads, 2ร—2 patch tokens (256 patches of four palette slots), partial axial 2D-RoPE, three palette slots and one style slot. See the README.
Parameters 66,424,320 (counted from the released model.safetensors; 224 float32 tensors)
Checkpoint SHA256 1a26888728f3c523075f691e4fddb270889e0a4db814c516da016709a06b9dbc (model.safetensors)
Output One 32ร—32 RGBA sprite per sample; index 0 is transparent
Text encoder External, not shipped: the first 11 layers of the openai/clip-vit-large-patch14 text model (penultimate features, 18 positions)
Intended use Generate palette-indexed pixel-art game sprites (items, weapons, characters, creatures, props) from a short English prompt.
Out of scope Photoreal images, text rendering, anything larger than 32ร—32, and subjects far outside game-sprite content.

Training data

Field Value
Training set About 130k native 32ร—32 sprites generated for this project with Krea 2 Turbo plus a pixel-art LoRA, cut from 3ร—3 sheets by data-pipeline/extract_v2.py. Not released.
License of the data The generated sprites are Krea 2 Outputs, owned by their generator per section 5.3 of the Krea 2 Community License. See NOTICE.
Holdouts excluded The Blind-50 and Eval-50 phrase lists were excluded from training by the builder's frozen-holdout check. The lists themselves are not published.
Personal data None in the weights files. Captions are the generation prompts.

Weights license

Field Value
License Apache-2.0, the same as the code. See LICENSE and NOTICE.
Inherited restrictions None claimed. No Krea or CLIP weights are part of the release; CLIP (MIT) is downloaded separately at sampling time.

Evaluation

Field Value
IID validation NLL 0.5237 at the released step (lowest of the run, which is why this checkpoint was kept)
Blind-50 / Eval-50 Not reported as a number for this release.

Known limitations

  • Stray magenta pixels: the training sheets used a magenta key background, and some samples keep a few magenta background-key pixels near the edges.
  • The run overfits after about 1.25M images; later checkpoints were worse on IID NLL, so this is not the last checkpoint of the run.
  • Needs the external CLIP ViT-L/14 text encoder, which is linked, not shipped.
  • Quality varies by subject; many samples are rough. Treat it as a small research model, not a finished asset generator.

How to sample

git clone https://github.com/QuarkOS/pixie && cd pixie
python3 -m venv .venv && . .venv/bin/activate
pip install torch numpy pillow tokenizers safetensors huggingface_hub
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

Each row of sprites.png is one prompt (32ร—32 sprites, upscaled with nearest-neighbour).

License

Apache-2.0 for the weights. See LICENSE and NOTICE (Krea 2 attribution and third-party credits).

Downloads last month
-
Safetensors
Model size
66.4M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support