speridlabs' single-image-to-3D model: Given an object image (rendered, generated, or photo), it automatically removes the background, crops the image, generates up to 1,048,576 3D Gaussians, and even reconstructs sides not visible in the original image, saving the result as a standard 3DGS .ply file. Originally included on 10 31; re-included this time because Comfy-Org released a ComfyUI package (Comfy-Org/Idol, uploaded by Kijai) on 11 1, though the model itself remains unchanged.
Rising score 491 · Likes 516 · Downloads 10742
Reasons
- New features compared to the version from 10 31: Comfy-Org's packaging notes state that Idol is a fine-tuned version of TripoSplat, allowing direct use with existing TripoSplat workflows; the package includes the DINOv3 image encoder (bf16), FLUX.2 VAE, and the Idol decoder, ready to use once placed in the correct directories.
- Official comparisons on H100 against TripoSplat, TRELLIS, and LGM: Input view SSIM 0.726, LPIPS 0.213, novel view DINOv2 0.781, and CLIP 0.885 scores are all the best in the table; single request time is 6.8 seconds (TRELLIS: 7.1 seconds, TripoSplat: 9.9 seconds).
- Runs on consumer-grade GPUs: Officially stated that the entire request involves approximately 1.80B parameters and peak VRAM usage of 7.2 GB; with 50 steps and 262,144 Gaussians, RTX 4090 takes about 19.1 seconds per image, while RTX 3080 takes about 46.2 seconds.
Usage
- ComfyUI: Download from Comfy-Org/Idol. Place idol_fp16.safetensors in models/diffusion_models, dino_v3_vit_h_bf16.safetensors in models/clip_vision, and flux2-vae.safetensors along with idol_decoder_fp16.safetensors in models/vae. Then open an existing TripoSplat workflow and replace the files accordingly.
- Command line: First accept the terms for facebook/dinov3-vith16plus-pretrain-lvd1689m and black-forest-labs/FLUX.2-dev on Hugging Face, run hf auth login, then git clone speridlabs/idol-sdk, pip install -e ., and execute python -m idol.sample --image photo.jpg --mode real (use 'real' for photos, 'synthetic' for rendered or generated images).
- Common parameters: --steps 50, --cfg_strength 3.0, and --num_gaussians between 32,768 and 1,048,576 (multiples of 32, default 262,144); open the resulting .ply file with any Gaussian splatting viewer.
Risks
- Evaluations are self-reported by the official team; LGM achieves the highest PSNR score (officially explained as PSNR favoring blurred results); this column has not reviewed demo videos.
- The Comfy-Org page only provides file placement instructions and does not include an Idol-specific workflow, requiring users to modify existing TripoSplat workflows themselves; the DINOv3 and FLUX.2 VAE in the package have their own license terms and do not follow Idol's Apache-2.0 license.
- The output is Gaussian splatting, not a mesh, so conversion is required for use in game engines or 3D printing; it is designed only for single objects.