Trout / README.md
FrameXlabs's picture
switch Space to static hosting (account at cpu-basic quota; gradio app included for one-click activation)
ff76bdd verified
|
Raw History Blame Contribute Delete
1.67 kB
---
title: Trout Pixel Studio
emoji: 🐟
colorFrom: indigo
colorTo: blue
sdk: static
pinned: false
license: apache-2.0
tags:
- pixel-art
- text-to-image
- masked-diffusion
- discrete-diffusion
- diffusion-lm
short_description: 16x16 pixel-art sprites from text (47.6M model)
---
# 🐟 Trout Pixel Studio
Static showcase for **[Trout](https://huggingface.co/FrameXlabs/Trout)** β€” a 47.6M-parameter
**absorbing-state masked diffusion language model** for 16Γ—16 pixel art, built and trained
**from scratch** (architecture after
[Tessera-small](https://huggingface.co/thejagstudio/tessera-small) by Jagrat Patel
(TheJagStudio), trained on [Pixel Art Atlas 1.4M](https://huggingface.co/datasets/Angais/pixel-art-atlas-1.4m),
CC0, conditioned on a frozen CLIP ViT-B/32 text tower).
Each sprite is a 256-token "sentence" over a learned **512-color palette** (token 0 =
transparent). A bidirectional hourglass transformer (width 512, 2+8+2 blocks, QK-RMSNorm,
2D axial RoPE, Conv-SwiGLU, adaLN-single, registers, self-conditioning) unmasks pixels
**in parallel over ~16 steps**. No VAE, no latent space, no autoregression.
**Why static?** Hugging Face currently requires a PRO subscription to host *Gradio*
Spaces on free cpu-basic hardware, so this Space is a free static page. A complete,
tested interactive **Gradio demo is included in this repo** (`app.py`, `trout.py`,
`requirements.txt`): switching this Space's SDK to `gradio` (with `app_file: app.py`)
makes the live demo run unchanged, or run it locally with `python app.py`.
**License:** Apache-2.0 (Trout weights + this page). The CLIP text tower used for
conditioning keeps its own MIT license.