title: README
emoji: 🧩
colorFrom: gray
colorTo: blue
sdk: static
pinned: false
🌐 Web page | 📄 Paper | 💻 GitHub
Generative models that run at any compute budget. BudgetAny collects open-source implementations of compute-adaptive ("budget") tokenization. A generative model is fine-tuned so that one checkpoint runs on a variable number of tokens, chosen at inference: detailed content keeps fine tokens, simple content is merged into coarse ones, and compute and memory fall with the token count.
Contributions are welcome. Bug reports, questions and pull requests on GitHub are all appreciated, from a typo fix or clearer documentation to a port to a new base model or a new modality. For larger changes, please open an issue first so we can plan it together.
Unofficial, open-source implementation. The method comes from BudgetPix: Compute-Adaptive Tokenization for Pixel-Space Image Diffusion (Kara et al., 2026). The code and models here are a community reimplementation and extension of it, not an official release, and BudgetAny is not a company, a lab or a formal organisation: the name only groups the repositories.
Categories
| Category | Task | Code | Page | Paper | |
|---|---|---|---|---|---|
 |
BudgetPix | pixel-space image generation | BudgetPix | budgetpix | arXiv:2610.12307 |
 |
Budget3D | image-to-3D generation (3D Gaussian splats) | Budget3D | budget3d | tech report coming soon |
 |
BudgetLatent | latent-space image generation | coming soon | budgetlatent | tech report coming soon |
Models
BudgetPix
| Model repository | Base model | Task | License of the weights |
|---|---|---|---|
| BudgetAny/BudgetPix-JiT | JiT B/16, B/32, L/16, L/32 | class-conditional ImageNet, 256 and 512 px | MIT |
| BudgetAny/BudgetPix-pMF | pixel MeanFlow L/16 | one-step class-conditional ImageNet, 256 px | MIT |
| BudgetAny/BudgetPix-MiniT2I | MiniT2I B/16, L/16 | text-to-image, 512 px | MIT |
| BudgetAny/BudgetPix-PixelDiT | PixelDiT-T2I 1.3B | text-to-image, 1024 px | NVIDIA License (NSCLv1), non-commercial |
Budget3D
| Model repository | Base model | Task | License of the weights |
|---|---|---|---|
| BudgetAny/Budget3D-Idol | speridlabs/idol | image-to-3D, 50 steps | Apache-2.0 |
| BudgetAny/Budget3D-Idol-Turbo | speridlabs/idol-turbo | image-to-3D, 5 steps | Apache-2.0 |
BudgetLatent
Coming soon.
Citation
If you use these models or the code, please cite the BudgetPix paper, and BudgetAny:
@article{kara2026budgetpix,
title={BudgetPix: Compute-Adaptive Tokenization for Pixel-Space Image Diffusion},
author={Ozgur Kara and Yujia Chen and Daniel Watson and David Forsyth and James Matthew Rehg and Wen-Sheng Chu and Du Tran},
journal={arXiv preprint arXiv:2610.12307},
year={2026}
}
@misc{budgetany2026,
title={BudgetAny: Generative Models at Any Compute Budget},
author={Ozgur Kara},
year={2026},
howpublished={\url{https://budgetany.github.io}}
}
Built on JiT, pixel MeanFlow, MiniT2I, PixelDiT, Idol and Idol-turbo, FLUX.2, DINOv3 and APT.
We acknowledge the compute support of the Health Care Engineering Systems Center within the Coordinated Science Lab at UIUC.