Overview
Idol is an open image-to-3D model from Speridlabs, released on October 28. Give it a photo or a render and it produces a complete 3D object represented as 3D Gaussian splats, inferring the sides the image never showed. Architecturally it is a 3D diffusion transformer trained from scratch, with about 659M parameters trained by Speridlabs: a 370.6M DiT, a 50.8M octree decoder and a 237.3M Gaussian decoder.
Inference is undemanding: peak VRAM stays under 7.2GB, so consumer GPUs are enough. Speridlabs reports 262,144 Gaussians at 50 denoising steps taking 6.7 seconds on an H100 and 19.1 seconds on an RTX 4090, with an RTX 5090 landing around 14 seconds. Gaussian count scales from 32,768 to 1,048,576, trading detail for file size.
Output is a .ply splat file plus the predicted camera parameters, with a .glb preview for ordinary 3D pipelines. Two input modes distinguish photographs from renders, changing only how much margin the crop leaves around the object. Weights, SDK and demo are Apache 2.0, and it also runs on fal in about 10 seconds without local setup.
Key Features
- Complete object from one image: Starting from a single photo, the model infers the back and sides and returns a rotatable 3D object.
- Gaussian splat output: Produces a standard .ply splat file usable in common 3D and visualisation pipelines, with the predicted camera attached.
- Adjustable density: Gaussian count ranges from 32,768 to 1,048,576 in multiples of 32, defaulting to 262,144 to balance detail against file size.
- Low VRAM: Peak memory under 7.2GB means 19 seconds on an RTX 4090 and about 14 on an RTX 5090, so local iteration needs no queue.
- Two input modes: real suits photographs and synthetic suits renders and generated images, differing in how much margin the crop leaves.
- Apache 2.0: Weights, SDK and demo are released under Apache 2.0, allowing commercial use and further training, with a hosted fal endpoint as an option.
Use Cases
- Product photos become rotatable 3D previews for interactive storefront detail pages.
- Games and animation draft concept assets to check silhouette before committing to modelling.
- Turntables and moodboards use lit 3D objects instead of flat images.
- Spatial computing and robotics projects generate synthetic scene material.
- Generated images feed straight into 3D, chaining behind an image model.
- Occasional users call the fal endpoint rather than setting up a local GPU.
Pros
- A single image yields a complete 3D object, so the starting cost is minimal.
- Peak VRAM under 7.2GB runs on consumer graphics cards.
- Adjustable splat density lets you trade detail against file size.
- Exports .ply and .glb, slotting into existing 3D pipelines without conversion.
- Apache 2.0 covers weights, code and demo alike.
- The fal endpoint returns results in about 10 seconds for machines without a local GPU.
Pricing
Open source under Apache 2.0, with weights around 3GB downloaded automatically on first run and the SDK installed via pip. The first run requires accepting the terms of two gated Hugging Face models, DINOv3 and FLUX.2-dev, of which only the VAE is downloaded. Peak VRAM stays under 7.2GB, and Speridlabs reports per-GPU timings at 262,144 Gaussians and 50 steps: 3.4 seconds on a B300, 6.7 on an H100, 14.1 on an RTX 5090, 19.1 on an RTX 4090 and 46.2 on an RTX 3080. The hosted fal endpoint bills per call with pricing on the fal pricing page, and a free web demo is available.
Summary
Turning one image into 3D used to mean either reconstruction, which needs multiple views, or generation, where the model is too large to run locally. Idol splits the difference with a small 3D diffusion transformer: 600M scale, 7GB of VRAM, a splat file ready for the pipeline within twenty seconds. What it produces is a plausible-looking object rather than a measured reconstruction, a distinction that rarely matters for concept assets and product previews.
Version History
- Idol listed: single-image 3D Gaussian splatting, open source (2026-10-31): Form at listing: an open image-to-3D model from Speridlabs, a from-scratch 3D diffusion transformer with about 659M trained parameters, comprising a 370.6M DiT, a 50.8M octree decoder and a 237.3M Gaussian decoder. A single photo produces a complete 3D object exported as a .ply splat with camera parameters, with Gaussian counts adjustable from 32,768 to 1,048,576 and a default of 262,144. Peak VRAM stays under 7.2GB, with 6.7 seconds on an H100 and 19.1 on an RTX 4090. Weights, SDK and demo are Apache 2.0, and a fal endpoint returns results in about 10 seconds.