Sperid Labs' Idol Turns a Single Photo Into a Million-Gaussian 3D Asset
Sperid Labs released Idol, an Apache 2.0 image-to-3D model that generates up to 1M Gaussian splats in seconds on consumer GPUs.
- Sperid Labs released Idol, an Apache 2.0 image-to-3D Gaussian splat model.
- 600M parameter 3D-DiT runs on consumer GPUs using 7.2 GB VRAM, outputs standard .ply files.
- Generates up to 1,048,576 Gaussians per object, 3.4 s on B300, 6.7 s on H100.
- Rectified-flow transformer conditioned on DINOv3 and FLUX.2 VAE features, plus octree and Gaussian decoders.
- Beats TripoSplat, TRELLIS, LGM on SSIM, LPIPS, IoU, DINOv2 and CLIP metrics.
- Full stack open: SDK, weights, demo Space, ComfyUI and MLX ports.
Sperid Labs has released Idol, an open image-to-3D model that converts a single object image into a 3D Gaussian splat. It infers unseen surfaces, generates up to 1,048,576 Gaussians, and runs within a reported 7.2 GB of peak VRAM.
The project publishes its weights and SDK under Apache 2.0. A hosted Hugging Face Space provides a browser demo, while the SDK repository supports local generation on consumer and data-center GPUs.
From one picture to a portable splat
Idol accepts a render, generated image, or photograph of a single object. Its preprocessing removes the background and frames the subject before the model produces a PLY file commonly used by 3D Gaussian splat viewers. A companion JSON file records the camera parameters inferred from the source image.
A Gaussian splat represents an object as a cloud of soft 3D primitives. Each Gaussian stores position, color, opacity, scale, and orientation. Compatible renderers can draw these primitives in real time and display the object from novel viewpoints.
The output count can scale from about 32,000 to 1,048,576 Gaussians without retraining the model. Higher counts preserve more detail but require more generation time, storage, and rendering work.
The model behind the splats
Idol contains roughly 1.8 billion parameters across its full pipeline, including image encoders and decoders. Its core is a 370 million-parameter diffusion transformer conditioned on image features from DINOv3 ViT-H+/16 and the FLUX.2 VAE.
The transformer processes 8,192 object-latent tokens and one camera token. An octree decoder places anchor points in 3D space, then a Gaussian decoder expands every anchor into 32 Gaussians. This two-stage design allows the same model to generate several output sizes.
Rectified flow trains the transformer to move samples along direct trajectories from noise toward the target data distribution. Idol uses that formulation with a published 50-step sampling schedule, reducing the number of denoising passes required for generation.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
