Budget3D-Idol

🌐 Project page  |  💻 Code  |  📄 Tech report (coming soon) · BudgetPix paper  |  🧩 BudgetAny

One checkpoint, any token budget. Budget3D-Idol is a fine-tune of speridlabs/idol, a single-image to 3D Gaussian splat generator. At inference you choose how many of Idol's 8,192 latent tokens the diffusion transformer runs on: the regions where the latent varies least, empty space first, are merged into coarser tokens, and sampling gets faster and lighter. The same weights serve every budget; Idol's image encoders and Gaussian decoders are used unchanged. The 5-step model is BudgetAny/Budget3D-Idol-Turbo.

Unofficial, open-source implementation. This model is part of BudgetAny, a community reimplementation and extension of the compute-adaptive tokenization idea of BudgetPix: Compute-Adaptive Tokenization for Pixel-Space Image Diffusion (Kara et al., 2026). It is not an official release of that paper, and BudgetAny is not a company or a lab. The paper covers pixel-space image diffusion only; applying the idea to image-to-3D is an extension made here, and the numbers on this card are measurements of this project, not results from the paper. Thanks to the BudgetPix authors for the method.

Files

file	base model	dense tokens	parameters	size	SHA-256
budget3d_idol.safetensors	Idol (50 sampling steps)	8,192	378 M	1.51 GB	8d87669ab02823d6…

The file holds the fine-tuned denoiser (Idol's DiT and the Budget3D modules) as fp32 EMA weights. Also in the repository: perf/ (the complete speed and memory reports) and assets/ (the clips below; assets/fast/: the same objects sampled with optimised inference). Idol's image encoders and Gaussian decoders are downloaded from speridlabs/idol by the code.

Usage

git clone https://github.com/BudgetAny/Budget3D && cd Budget3D    # setup: see README.md
hf download BudgetAny/Budget3D-Idol budget3d_idol.safetensors --local-dir weights
python scripts/sample.py --ckpt weights/budget3d_idol.safetensors --image photo.jpg --budget 0.5 --out photo.ply

--budget is the fraction of the 8,192 tokens (0.0625 to 1); --fast switches on the optimised inference.

Speed and memory

One NVIDIA H100 80GB HBM3, 50 Euler steps with classifier-free guidance 3, batches of 8 objects, optimised inference (--fast): bf16 transformer blocks, TF32 elsewhere, cuDNN fused attention, torch.compile (blocks + adapters). Times are per object, averaged over the 72 objects in full batches after each GPU's warm-up batch; the speed-up is the sampling time at 100 % divided by the sampling time at the budget, for the same batch.

budget	tokens	sampling / object	speed-up	peak sampling memory	end to end / object
100 %	8,192	5.67 s	1.00×	9.9 GB	11.80 s
90 %	7,373	5.15 s	1.10×	9.6 GB	10.85 s
80 %	6,554	4.59 s	1.24×	9.2 GB	10.34 s
70 %	5,734	4.10 s	1.39×	8.9 GB	9.64 s
60 %	4,915	3.60 s	1.58×	8.5 GB	9.33 s
50 %	4,096	3.18 s	1.79×	8.2 GB	8.96 s
40 %	3,277	2.66 s	2.14×	7.8 GB	8.62 s

Peak memory includes about 4.5 GB of resident weights and image encoders. End to end adds decoding the Gaussians (2.53 s) and rendering a 72-frame orbit (3.60 s), which do not depend on the budget. The optimised settings speed up the stock Idol in the same way; the budget speed-ups come on top of them. This run covers budgets from 100 % to 40 %; the complete reports are in perf/.

The same in full precision (fp32), down to 10 % of the tokens

budget	tokens	sampling / object	speed-up	peak sampling memory	end to end / object
100 %	8,192	63.44 s	1.00×	14.8 GB	67.60 s
90 %	7,373	56.25 s	1.13×	14.1 GB	60.33 s
80 %	6,554	49.58 s	1.28×	13.4 GB	53.70 s
70 %	5,734	43.28 s	1.47×	12.8 GB	47.37 s
60 %	4,915	37.48 s	1.69×	12.1 GB	41.57 s
50 %	4,096	32.20 s	1.97×	11.4 GB	36.25 s
40 %	3,277	27.15 s	2.34×	10.7 GB	31.22 s
30 %	2,458	22.50 s	2.82×	10.1 GB	26.66 s
20 %	1,638	18.28 s	3.47×	9.4 GB	22.53 s
10 %	819	14.48 s	4.38×	8.7 GB	18.82 s

Averaged over the 160 objects in full batches after each GPU's warm-up batch.

Examples

Each clip shows the input image and the same object generated at 100 % down to 40 % of the tokens, from the same seed and camera orbit, with the sampling speed-up and peak memory of each budget (full precision). The objects were picked by hand from the Idol repository's example images.

Citation

Tech report coming soon. Until then, please cite the BudgetPix paper:

@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}
}

Built on Idol by Sperid Labs, DINOv3, the FLUX.2 VAE and Dome-Objaverse renders of Objaverse.

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

License

Apache-2.0. The weights are a fine-tune of speridlabs/idol, which is released under Apache-2.0. Running the model also needs DINOv3 ViT-H+/16 and the FLUX.2 VAE, which are not part of this repository and have their own licenses.