Tessera-small

Tessera-small

Generates 16×16 pixel art sprites from a text prompt with a small diffusion model that fills in all 256 pixels in about 16 steps. For non-commercial game and art experiments.

Made by thejagstudio on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Tessera-small is a text-to-image model for tiny pixel art, with 47.6M parameters. It treats a 16×16 sprite as 256 tokens, each one a color from a 512-color palette. Starting from a fully masked canvas, it fills in every pixel at once over about 16 steps, guided by your text.

Who it suits: It suits hobbyists, students and researchers who want small pixel art sprites for non-commercial projects, or who want to study a compact diffusion model.

Industry | Who uses it | How they use it | Example
Game prototyping | Hobbyist game makers | Generate placeholder sprites for a personal project | A hobbyist makes eight sprites for a personal game prototype
Education | Teachers and students | Study masked diffusion on a tiny image size | A class tries different step counts on the same prompt
Research | Researchers on discrete diffusion | Inspect the checkpoint, configs and training log | A researcher reads metrics.jsonl to follow training

Anyone who plans to sell the images or use them in a commercial product can skip it, because the license rules that out.

Where to use it: It runs from Python with PyTorch, and the card's snippet targets a CUDA GPU.

• Windows, macOS or Linux with Python: download the checkpoint with the Hugging Face tools
• A CUDA GPU: the device used in the card's Python example
• The Tessera Pixel Studio demo Space: a hosted demo that also holds the pipeline code

Use case | Who | What they get back
Make sprite ideas | Game hobbyists | A grid of 16×16 sprites for each prompt
Fill in part of a sprite | Pixel artists | An inpainted result, since the sampler supports inpainting
Study a small diffusion model | Students | Config files and a training log to read

How to set it up: You need a few things first.

• Python with PyTorch and the huggingface_hub package
• The tessera pipeline code, which the card says to take from the files of the demo Space
• The frozen CLIP text encoder openai/clip-vit-base-patch32, which is not included
• A CUDA GPU if you copy the card's snippet as it is

1. Install the Hugging Face hub package.
2. Download the model files.
3. Get the tessera pipeline code from the demo Space's files.
4. Run the Python snippet from How to use it.

pip install -U huggingface_hub
hf download thejagstudio/tessera-small

How to use it: The snippet loads the checkpoint, builds the pipeline on a CUDA device, asks for eight sprites per prompt over 16 steps, and saves them as one grid image.

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")

Tool | What it does
hf_hub_download | Fetches one file from the model repository, here the fp16 checkpoint
TesseraPipeline.from_checkpoint | Builds the generation pipeline from the checkpoint on the device you name
pipe(...) | Takes a list of prompts plus the sample count, steps, guidance scale and seed
pipe.grid(...).save(...) | Arranges the sprites into one grid image and writes it to a file

Example: You change the prompt in the snippet to "red mushroom with white spots" and run it. The pipeline samples eight 16×16 sprites for that prompt in 16 steps, and the grid image is saved under the file name you set.

For newcomers and experts: Tessera draws tiny pixel art from a short phrase. Your first small step is to try the snippet with one prompt and look at the saved grid, or to open the Tessera Pixel Studio demo that the card names.

Under the hood it is a bidirectional transformer that predicts masked palette tokens directly. Nothing is compressed into a latent space, and it does not generate pixel by pixel. The sampler unmasks pixels in parallel by confidence, uses classifier-free guidance, and supports inpainting. The trade-offs are a fixed 16×16 size, a 512-color palette and a non-commercial license.

Your prompt goes to a frozen CLIP text encoder
A fully masked canvas of 256 tokens starts
All pixels are unmasked in parallel over about 16 steps
Palette tokens turn into a pixel art image

Good to know:

• Token 0 in the palette is transparent, so sprites can have see-through areas.
• The model was trained on the Pixel Art Atlas 1.4M dataset by Angais, for about 5.5 hours on one NVIDIA RTX PRO 6000 Blackwell.
• On validation data it reached 78.8% accuracy on masked pixels, with a perplexity of 2.36 out of 512.

Title: CLIP score on ViT-B/32
Generated sprites: 25.24 points
Real sprites: 25.28 points

• License: CC BY-NC 4.0. Education, research and personal projects are fine with credit to the author, and anything commercial needs a separate license from the author. The CLIP text encoder and the training dataset keep their own licenses.
• Cost to run: none named, no API key, account or paid service is mentioned.
• Data it reaches: it downloads the checkpoint from Hugging Face and writes the image files you name.
• Not for: commercial use of any kind, including selling the model or its images, or using them in a paid product or a hosted service that earns money.

On Hugging Face, 13 November 2026

• Task: text to image
• Library: pytorch
• Downloads in the last 30 days: 0
• Likes: 3
• License: cc-by-nc-4.0
• Tags: pytorch, pixel-art, text-to-image, masked-diffusion, discrete-diffusion, diffusion-llm

Links

• Model: https://huggingface.co/thejagstudio/tessera-small
• Files: https://huggingface.co/thejagstudio/tessera-small/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download thejagstudio/tessera-small

Then follow the usage in its model card. Nothing in this kit runs by itself.