Original title: idol-turbo
Idol takes a single object image as input and outputs a 3D Gaussian point cloud of the entire object (including sides not visible in the image). Idol Turbo uses a two-stage LoRA (one for the high-noise stage and one for the low-noise stage) to distill the original 50-step process with CFG into 5 steps without requiring CFG. The authors claim it takes approximately 0.75 seconds on an H200; the base model remains unchanged.
Why it matters Image-to-3D Gaussian generation is reduced from the original multi-step process with CFG to just 5 steps, and it can be used in ComfyUI by adding two LoRAs.
Hugs for likes 38 · Hugs for upvotes 37 · Hugs for downloads 391
License: Commercial use allowed · ComfyUI: Mentioned in the model card · VRAM: Not specified
Reasons
- The authors state that the LoRA key names match the Comfy-Org/Idol package and have been pre-scaled to a strength of 1.0, allowing direct use with the official TripoSplat workflow
- The model card includes comparison images of the original and Turbo versions, both running at 5 steps, with one object not included in the training set
Usage
- First, set up the base model, DINOv3, flux2-vae, and decoder according to Comfy-Org/Idol. Then, place the slow and fast LoRAs from the loras/ directory into models/loras with a strength of 1.0
Risks
- The speed figures were measured by the authors on an H200; the distillation training set consists of single-object images selected by the authors, and complex scenes are not addressed