Budget3D: Image-to-3D Generation at Any Token Budget

Code Models Tech report: coming soon BudgetPix paper BibTeX

BudgetPix's compute-adaptive tokenization, adapted to image-to-3D generation (Idol, Idol-Turbo).

Test-time compute adjustment. Budget3D sets the token budget of Idol and Idol-Turbo at inference time, from 100% down to 40% of the 8,192 latent tokens, with the sampling time and peak memory at each budget.

Overview

Budget3D fine-tunes Idol and Idol-Turbo, which turn one image into 3D Gaussian splats, so that the number of tokens is a setting you choose at inference. The regions where the latent varies least, empty space first, are merged into coarser tokens, and the full-resolution tokens go where it varies most.

Budget3D Across Budgets

Twelve Idol and twelve Idol-Turbo objects as the budget goes from 100% to 40%.

Idol and Idol-Turbo

Budget3D-Idol (50 steps)

Budget3D-Idol-Turbo (5 steps)

Speed and Memory

Sampling time per object and peak memory, with the optimised inference of the code (bf16 blocks, fused attention, torch.compile) at every budget, 100% included; decoding to Gaussians and rendering add about 3 s per object at every budget.

Budget3D-Idol, one H100, batches of 8

Budget	Tokens	Sampling / object	Speed-up	Peak memory
100%	8,192	5.67 s	1.00×	9.9 GB
90%	7,373	5.15 s	1.10×	9.6 GB
80%	6,554	4.59 s	1.24×	9.2 GB
70%	5,734	4.10 s	1.39×	8.9 GB
60%	4,915	3.60 s	1.58×	8.5 GB
50%	4,096	3.18 s	1.79×	8.2 GB
40%	3,277	2.66 s	2.14×	7.8 GB

Budget3D-Idol-Turbo, one RTX PRO 6000, batches of 8

Budget	Tokens	Sampling / object	Speed-up	Peak memory
100%	8,192	0.48 s	1.00×	8.0 GB
90%	7,373	0.44 s	1.09×	7.8 GB
80%	6,554	0.39 s	1.22×	7.5 GB
70%	5,734	0.35 s	1.36×	7.3 GB
60%	4,915	0.31 s	1.52×	7.1 GB
50%	4,096	0.27 s	1.74×	6.9 GB
40%	3,277	0.25 s	1.95×	6.6 GB

BibTeX

@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}}
}

Acknowledgements

Thanks to Sperid Labs for releasing Idol and Idol-Turbo, which Budget3D builds on.

We acknowledge the compute support of the Health Care Engineering Systems Center within the Coordinated Science Lab at UIUC.