Budget3D-Idol

Budget3D-Idol is a fine-tuning of the single-image-to-3D Gaussian Splatting model Idol by the community project BudgetAny: during inference, you can specify using only a subset of Idol's 8192 latent tokens (6.25%–100%), where regions with the least variation (starting with blank areas) are merged into coarser tokens. The authors measured on H100 that at a 50% budget, the sampling time per object drops from 5.67 seconds to 3.18 seconds, and peak VRAM usage decreases from 9.9 GB to 8.2 GB; the related Budget3D-Idol-Turbo requires only 5 sampling steps. The image encoder and Gaussian decoder remain the same as the original Idol, with an online demo and community WebUI available.

Budget3D-Idol is a fine-tuning of the single-image-to-3D Gaussian Splatting model Idol by the community project BudgetAny: during inference, you can specify using only a subset of Idol's 8192 latent tokens (6.25%–100%), where regions with the least variation (starting with blank areas) are merged into coarser tokens. The authors measured on H100 that at a 50% budget, the sampling time per object drops from 5.67 seconds to 3.18 seconds, and peak VRAM usage decreases from 9.9 GB to 8.2 GB; the related Budget3D-Idol-Turbo requires only 5 sampling steps. The image encoder and Gaussian decoder remain the same as the original Idol, with an online demo and community WebUI available.

Why it matters The same weights can compute only a portion of the tokens on demand, merging blank areas first; lower budgets result in faster sampling and reduced VRAM usage. There is also a Turbo version with fewer steps.

Hugging Face likes 8 · Hugging Face trending score 8 · Hugging Face downloads 23

License: Commercial use allowed · ComfyUI: Not available · VRAM: ≈2 GB (estimated)

Reasons

  • Apache-2.0 license; the model card includes complete speed and VRAM reports, noting that these figures are self-tested by this project and not results from the paper
  • An online demo space is available on Hugging Face, allowing you to try it without installation

Usage

  • After running git clone https://github.com/BudgetAny/Budget3D,下载 to get budget3d_idol.safetensors and config.json, run python scripts/sample.py --ckpt weights/budget3d_idol.safetensors --image photo.jpg --budget 0.5 --out photo.ply; add --fast to enable optimized inference

Risks

  • This is an unofficial implementation; it is unrelated to the authors of the BudgetAny and BudgetPix papers or the Idol authors; the technical report has not yet been released
  • Running it requires separately downloading DINOv3 and FLUX.2 VAE, each with their own licenses; ComfyUI support is currently unavailable

Open

抱脸 image-to-3d 上升榜

Explore the latest AI Radar updates