speridlabs/idol

Speridlabs' single-image-to-3D model: Input a photo, render, or generated image of an object; it automatically removes the background, crops the composition, and then generates up to 1,048,576 3D Gaussians (including sides and backs not captured in the image), saving them as a standard 3DGS .ply file that can be opened directly with a Gaussian splatting viewer.

2026-10-31T01:38:31Z

Rising Score 149 · Likes 150 · Downloads 221

Reasons

Low barrier to entry: The entire inference process involves approximately 1.80B parameters (370.6M denoising DiT + octree decoder + Gaussian decoder, plus BiRefNet for background removal, DINOv3 encoder, and FLUX.2 VAE); the authors state peak VRAM usage is 7.2 GB; 50 steps and 262,144 Gaussians take 19.1 seconds on an RTX 4090, 46.2 seconds on an RTX 3080, and 3.4 seconds on a B300.
Authors' self-tests compared with TripoSplat, TRELLIS, and LGM: For the input view, SSIM is 0.726, LPIPS is 0.213, and IoU is 0.862; for seven novel views, DINOv2 0.781 and CLIP 0.885 scores are the best among the four models; PSNR is 12.25, lower than LGM's 12.46, which the authors explain is because LGM produces blurrier images but higher PSNR scores.
Released on 28 10; the code (idol-sdk), weights, and online demo Space are all public under the Apache-2.0 license; this was not included in previous issues. The Rising Score of 149 only reflects attention level.

Usage

First, agree to the terms for the gated models facebook/dinov3-vith16plus-pretrain-lvd1689m and black-forest-labs/FLUX.2-dev (downloading only its VAE), then log in using hf auth login; next, run git clone speridlabs/idol-sdk and pip install -e . (requires Python 3.10+, PyTorch 2.5+, and an NVIDIA CUDA GPU); the first run automatically downloads approximately 3 GB of weights.
Run python -m idol.sample --image photo.jpg --mode real (for photos where the object occupies 1/1.7 of the frame); for renders and generated images, use --mode synthetic (default, occupying 1/1.2 of the frame); --image can accept a folder for batch processing; each image outputs a .ply file, the cropped image actually seen by the model (_input.png), and the predicted camera parameters (_camera.json).
Common parameters: --steps 50 (fewer steps are faster but lose detail), --cfg_strength 3.0, --rescale_t 3.0 (higher values favor overall shape, lower values favor detail), --num_gaussians defaults to 262144 and can be set between 32,768 and 1,048,576 (multiples of 32); in Python, use IdolPipeline.from_pretrained("speridlabs/idol").

Risks

Only processes single objects; the output is a Gaussian splatting .ply point cloud, not a mesh, and the card does not specify methods for converting to mesh or exporting textures.
Depends on two gated models; FLUX.2-dev has its own license terms from Black Forest Labs, which must be reviewed before commercial use; only NVIDIA CUDA GPUs are supported.
Evaluations and speed tests are all self-reported by the authors (evaluations conducted on H100), with no paper or technical report; "within 2 seconds" is the authors' fastest claim, while the table shows the fastest default setting takes 3.4 seconds; this column has not reviewed example videos.

Open