Tessera-small: a diffusion language model for 16×16 pixel art

Author: Jagrat Patel (TheJagStudio) Live demo: Tessera Pixel Studio License: CC BY-NC 4.0. Free for education, research and personal non-commercial use with credit. Commercial use is not allowed without a separate license from the author.

Tessera treats a 16×16 sprite as a 256-token sentence. Every pixel is one token from a learned 512-color palette (token 0 = transparent). A bidirectional transformer starts from a fully masked canvas and unmasks all 256 pixels in parallel in about 16 steps, conditioned on a frozen CLIP text encoder. It has no VAE, no latent space and no autoregression.

Model


Parameters	47.6M (Tessera-small)
Architecture	joint text+pixel bidirectional transformer, hourglass 2 + 8 + 2 blocks (2×2 patch-merged core), width 512, 8 heads, QK-RMSNorm, 2D axial RoPE, Conv-SwiGLU, adaLN-single, 4 registers, self-conditioning
Text encoder	frozen `openai/clip-vit-base-patch32` text tower, 24 tokens (not included, MIT licensed)
Vocabulary	512 palette tokens (frequency-weighted OKLab k-means, ΔE 0.010) + `[MASK]`
Objective	absorbing-state masked diffusion, cosine schedule, blob masks, perceptual label smoothing, CFG dropout
Sampler	parallel confidence-ordered unmasking with CFG, Gumbel-annealed choice, spatial NMS; inpainting supported

Training

82,342 steps × 512 sprites (30 epochs over 1.4M sprites) on one NVIDIA RTX PRO 6000 Blackwell (≈5.5 h). The optimizer is Muon for hidden matrices plus AdamW, with a Warmup-Stable-Decay schedule, EMA 0.9995, bf16 and torch.compile.

Metric (EMA weights)	Value
Validation NLL	0.859 nats / pixel (perplexity 2.36 of 512)
Validation accuracy	78.8% of masked pixels
CLIP score (ViT-B/32)	25.24 generated vs 25.28 real sprites

Usage (non-commercial)

from huggingface_hub import hf_hub_download
from tessera.pipeline import TesseraPipeline   # code: see the demo Space's files

ckpt = hf_hub_download("thejagstudio/tessera-small", "tessera-small-fp16.pt")
pipe = TesseraPipeline.from_checkpoint(ckpt, device="cuda")
images = pipe(["golden crown with blue gems"], num_per_prompt=8, steps=16, cfg_scale=3.0, seed=0)
pipe.grid(images).save("crowns.png")

Files

File	Contents
`tessera-small-fp16.pt`	EMA weights (fp16), model config, palette, text-encoder id, training step
`model_config.json`, `train_config.json`	architecture and training hyperparameters
`metrics.jsonl`	full training and validation log
`showcase.png`	sample grid (16 prompts × 8 samples)

Credits

Huge thanks to Angais for creating and openly releasing Pixel Art Atlas 1.4M. This model was trained on it and would not exist without it.

License

Tessera is released under Creative Commons Attribution-NonCommercial 4.0.

✅ Allowed (free): education, teaching, academic and independent research, personal projects and experiments. You can download, run, fine-tune and share the model, with credit ("Tessera by Jagrat Patel (TheJagStudio)").
❌ Not allowed: any commercial use of the model or of images generated with it, such as selling or licensing it, paid apps, games or products, or monetized hosted services.
💼 Commercial license: contact Jagrat Patel via Hugging Face.

Third-party parts keep their own licenses: the CLIP text encoder (OpenAI, MIT) and the Pixel Art Atlas 1.4M dataset (Angais).

Downloads last month: -; Downloads are not tracked for this model. How to track

thejagstudio
/

tessera-small