Ozgur Kara (BudgetAny project) · Adaptive-budget image-to-3D
Budget3D
Budget3D fine-tunes Idol and Idol-Turbo, which turn a single image into 3D Gaussian splats, so that the number of latent tokens becomes an inference-time setting. Low-variance regions, empty space first, are merged into coarser tokens, while full-resolution tokens are kept where the latent varies most. It applies the compute-adaptive tokenization from BudgetPix. Reducing the budget from 100% to 40% of the 8,192 tokens cuts sampling time per object from 5.67 s to 2.66 s (2.14x) for Budget3D-Idol on one H100, and from 0.48 s to 0.25 s (1.95x) for Budget3D-Idol-Turbo on one RTX PRO 6000. Peak memory falls from 9.9 GB to 7.8 GB and from 8.0 GB to 6.6 GB respectively. Decoding and rendering add about 3 s per object. Code and models are linked from the page; the tech report is listed as coming soon, and no licence is stated.
Open the source ↗Similar 3D & Spatial releases
- LucidLyra
NVIDIA Research (Oliver Hahn, Rodrigo Marcuzzi, Sherwin Bahmani, Jiahui Huang, Laura Leal-Taixé and others) · Queryable 3D scenes · Week 46, 2026
- Idol
Speridlabs (Alejandro Cañada, Chema Garabito) · Image-to-3D splats · Week 45, 2026
- JST-2
Viggle AI · World Model · Week 44, 2026
- MoVerse
Orange 3DV Team · 3D Generation · Week 24, 2026
- Viggle AI
Viggle Inc. · 3D Generation · Week 20, 2026
- Viggle PINOC
Viggle AI · 3D Generation · Week 20, 2026
Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.