Budget3D-Idol-Turbo

Budget3D-Idol-Turbo is a fine-tune by the community project BudgetAny of Sperid Labs' Idol-turbo (a 5-step, guidance-free single-image-to-3D Gaussian model). It adapts the "compute-adaptive tokenization" concept from the BudgetPix paper from image diffusion to 3D: during inference, use --budget to select the proportion of the 8192 latent tokens that actually participate in computation, with the same weights adapting to all budgets. The authors claim that on an RTX PRO 6000, sampling is 1.74x faster at 50% budget and 3.22x faster at 10% budget, but end-to-end time per object remains around 3.2–3.6 seconds because Gaussian decoding and rendering are unaffected by the budget.

Budget3D-Idol-Turbo is a fine-tune by the community project BudgetAny of Sperid Labs' Idol-turbo (a 5-step, guidance-free single-image-to-3D Gaussian model). It adapts the "compute-adaptive tokenization" concept from the BudgetPix paper from image diffusion to 3D: during inference, use --budget to select the proportion of the 8192 latent tokens that actually participate in computation, with the same weights adapting to all budgets. The authors claim that on an RTX PRO 6000, sampling is 1.74x faster at 50% budget and 3.22x faster at 10% budget, but end-to-end time per object remains around 3.2–3.6 seconds because Gaussian decoding and rendering are unaffected by the budget.

Why it matters By merging blank and low-variance regions into coarse tokens based on budget, lower budgets result in faster sampling; the authors provide speed and VRAM tables for each tier, with peak VRAM at full budget around 8 GB.

Hugging Face likes 4 · Hugging Face upvotes 3 · Hugging Face downloads 291

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

Reasons

  • Apache-2.0; includes a Hugging Face online demo (50-step version) and a community WebUI
  • Examples in the model card are taken from the test set at fixed intervals, not cherry-picked

Usage

  • git clone https://github.com/BudgetAny/Budget3D and configure the environment according to its instructions, run hf download BudgetAny/Budget3D-Idol-Turbo budget3d_idol_turbo.safetensors config.json --local-dir weights, then run python scripts/sample.py --ckpt weights/budget3d_idol_turbo.safetensors --turbo --image photo.jpg --budget 0.5 --out photo.ply

Risks

  • Not the official implementation of the paper; the technical report has not been released yet, and all figures are self-tested by the authors
  • Running it also requires DINOv3 and FLUX.2 VAE, each with their own licenses; no ComfyUI nodes are available yet

Open

抱脸 image-to-3d 上升榜

Explore the latest AI Radar updates