Budget3D-Idol-Turbo

🌐 Project page  |  💻 Code  |  🎮 Demo (Budget3D-Idol)  |  🖥️ WebUI  |  📄 Tech report (coming soon) · BudgetPix paper  |  🧩 BudgetAny

One checkpoint, any token budget, 5 sampling steps. Budget3D-Idol-Turbo is a fine-tune of speridlabs/idol-turbo, the 5-step guidance-free distillation of the single-image to 3D Gaussian splat generator Idol. At inference you choose how many of Idol's 8,192 latent tokens the 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 50-step model is BudgetAny/Budget3D-Idol.

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_turbo.safetensors	Idol-turbo (5 sampling steps)	8,192	407 M	1.63 GB	4f0c31ed6979f804…

The file holds the fine-tuned denoiser (Idol's DiT, Idol-turbo's two LoRA experts and the Budget3D modules) as fp32 EMA weights. Also in the repository: perf/ (the complete speed and memory reports) and assets/ (the clips below, and the same objects from the stock Idol-turbo at 100 % in assets/stock_idol_turbo/). 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-Turbo budget3d_idol_turbo.safetensors config.json --local-dir weights
python scripts/sample.py --ckpt weights/budget3d_idol_turbo.safetensors --turbo --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.

For a graphical interface, try the online demo on Hugging Face Spaces (Budget3D-Idol, nothing to install) or run the community WebUI on your own GPU.

Speed and memory

One NVIDIA RTX PRO 6000 Blackwell Server Edition, 5 Euler steps without guidance, batches of 8 objects, optimised inference (--fast): bf16 transformer blocks, TF32 elsewhere, cuDNN fused attention, torch.compile. Times are per object, averaged over the 160 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	0.48 s	1.00×	8.0 GB	3.64 s
90 %	7,373	0.44 s	1.09×	7.8 GB	3.51 s
80 %	6,554	0.39 s	1.22×	7.5 GB	3.45 s
70 %	5,734	0.35 s	1.36×	7.3 GB	3.37 s
60 %	4,915	0.31 s	1.52×	7.1 GB	3.25 s
50 %	4,096	0.27 s	1.74×	6.9 GB	3.27 s
40 %	3,277	0.25 s	1.95×	6.6 GB	3.39 s
30 %	2,458	0.22 s	2.21×	6.4 GB	3.21 s
20 %	1,638	0.17 s	2.74×	6.2 GB	3.15 s
10 %	819	0.15 s	3.22×	6.0 GB	3.23 s

Peak memory includes about 4.6 GB of resident weights and encoders. End to end adds decoding the Gaussians (0.92 s) and rendering a 72-frame orbit (2.25 s), which do not depend on the budget. The speed-up is smaller than the token saving because every sequence also carries 4,102 image-condition tokens that the budget does not shorten. At 100 % the model costs the same as the stock Idol-turbo in the same setting (0.47 s per object, median). The complete reports are in perf/.

Examples

Each clip shows the input image and the same object generated at 100 % down to 10 % of the tokens, from the same noise and camera orbit. The objects are not picked by quality: they are every 51st of the 200 test objects (2 example images from the Idol repository and 2 held-out Dome-Objaverse objects).

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-turbo and 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-turbo, which is released under Apache-2.0 and is itself derived from speridlabs/idol (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.