✦
Summary
Budget3D is an open-source image-to-3D system that turns a single image into 3D Gaussian splats, with a token budget the user sets at inference time to trade quality for speed and memory. It fine-tunes the Idol and Idol-Turbo image-to-3D models so one checkpoint runs at any budget from 40% to 100% of 8,192 tokens, merging low-variation latent regions into coarser tokens when the budget drops. The code is MIT-licensed and the model weights are Apache-2.0, and a free Hugging Face Spaces demo lets anyone try it in the browser with nothing to install.
Key Features
Compute-adaptive tokenization with a token budget adjustable from 100% down to 40% at inference
One checkpoint runs at any token budget without retraining
Turns a single image into 3D Gaussian splats exported as .ply files
Two models: Budget3D-Idol with 50 steps and Budget3D-Idol-Turbo with 5 steps
Up to about 2.14x faster sampling at 40% budget with lower peak memory
Free in-browser demo on Hugging Face Spaces with nothing to install
Local Gradio WebUI with a token-budget slider, interactive 3D view and turntable video
Optimised inference with bf16 blocks, fused attention and torch.compile under MIT and Apache-2.0 licences
About Budget3D
At the lowest 40% budget, sampling runs up to about 2.14 times faster with lower peak memory: the full Idol model drops from 5.67 to 2.66 seconds per object, and the distilled Idol-Turbo from 0.48 to 0.25 seconds. Two models ship in the repo, Budget3D-Idol with 50 sampling steps and Budget3D-Idol-Turbo with just 5 steps, and both output standard .ply Gaussian splat files. Local use needs Python 3.11 to 3.13 and a CUDA GPU, via a command-line sampler or a local Gradio WebUI with a token-budget slider, an interactive 3D view, turntable video export and .ply download.
Budget3D was released in November 2026 by BudgetAny and is aimed at 3D artists, game developers and researchers who need fast single-image 3D reconstruction on limited hardware. Inference is optimised with bf16 blocks, fused attention and torch.compile, and the weights are published on Hugging Face as BudgetAny/Budget3D-Idol and BudgetAny/Budget3D-Idol-Turbo. Because the budget is chosen per generation, the same checkpoint serves both quick previews and full-quality reconstructions.