Tessera-small: A small model that generates pixel art as sentences

Uploaded by individual developer Jagrat Patel (TheJagStudio) on 11 9. Tessera treats each pixel of a 16×16 sprite as a token (selected from a learned 512-color palette, where index 0 is transparent). A bidirectional Transformer starts from a fully masked canvas, conditioned on a frozen CLIP text encoder, and reveals all 256 pixels in parallel in approximately 16 steps; there is no VAE, no latent space, and it is not autoregressive. Tessera-small has 47.6M parameters, was trained on the Pixel Art Atlas 1.4M dataset, and is licensed under CC BY-NC 4.0.

Original title: thejagstudio/tessera-small

Uploaded by individual developer Jagrat Patel (TheJagStudio) on 11 9. Tessera treats each pixel of a 16×16 sprite as a token (selected from a learned 512-color palette, where index 0 is transparent). A bidirectional Transformer starts from a fully masked canvas, conditioned on a frozen CLIP text encoder, and reveals all 256 pixels in parallel in approximately 16 steps; there is no VAE, no latent space, and it is not autoregressive. Tessera-small has 47.6M parameters, was trained on the Pixel Art Atlas 1.4M dataset, and is licensed under CC BY-NC 4.0.

Why it matters With fewer than fifty million parameters and training in just a few hours on a single GPU, it is suitable for game icons and pixel assets, and its architecture is also good for learning; non-commercial use only

Rising score 46 · Likes 51 · Downloads 0

Reasons

  • The author states it was trained for approximately 5.5 hours on a single RTX PRO 6000 Blackwell (82,342 steps, 30 epochs)
  • Prediction accuracy for masked pixels on the validation set is 78.8%; CLIP score is 25.24, compared to 25.28 for real sprites
  • Supports inpainting; an online demo Space, Tessera Pixel Studio, is available

Usage

  • Use hf_hub_download to download tessera-small-fp16.pt and load it with TesseraPipeline.from_checkpoint (code is in the files of the demo Space); example parameters are steps=16, cfg_scale=3.0, num_per_prompt=8
  • The text encoder from openai/clip-vit-base-patch32 is also required (not included, MIT license); text is limited to 24 tokens

Risks

  • CC BY-NC 4.0: Neither the model nor the generated images can be used commercially (cannot be sold or included in paid games or apps); commercial use requires a separate agreement with the author
  • Only outputs small 16×16 images; inference code is not in the model repository and must be obtained from the demo Space. This entry has not been run

Open

Hugging Face Trending

Explore the latest AI Radar updates