Idol transforms your creations into 3D objects with an open model

Speridlabs introduces Idol, an open-source model capable of turning an image into a 3D object, with 7.2 GB of VRAM and multiple levels of detail.

A photograph, an illustration, or an image generated by artificial intelligence can now become a three-dimensional object observable from all angles, without going through a manual modeling step. This is the proposition of Idol, a new model developed by Speridlabs, capable of reconstructing a 3D object from a single image. Its relatively compact architecture, its claimed performance, and the release of its weights under the Apache 2.0 license make it a new contribution to the open 3D generation ecosystem.

Developed by Alejandro Cañada and Chema Garabito, Idol relies on a Diffusion Transformer architecture, or 3D-DiT, trained to directly produce three-dimensional representations using the 3D Gaussian Splatting technique. The company provides an interactive demo, inference code on GitHub, and model weights on Hugging Face. Reconstructing the unseen sides of an object

The difficulty of 3D generation from a single image stems from the lack of information about hidden parts. A photograph shows the front face of an object, but provides no direct information about its back, its thickness, or certain proportions.

To produce a complete representation, Idol relies on knowledge acquired during its training and on the visual features of the provided image. The model thus estimates the overall shape of the object, its depth, and the appearance of its surfaces, including those that do not appear in the original image.

The examples published by Speridlabs cover a wide variety of objects: armchairs, robots, electronic devices, shoes, furniture, food, or stylized characters. Each generation can be manipulated in a 3D viewer to examine the different sides of the reconstructed object.

This reconstruction nevertheless carries an important limitation. The unseen parts are not retrieved from physical measurements: they are estimated by the model. The resulting appearance may be consistent without matching that of the photographed object exactly. Up to one million Gaussians to represent an object

Unlike traditional 3D modeling, which generally represents surfaces using meshes made of vertices and faces, Idol uses 3D Gaussian Splatting.

This technique relies on a set of three-dimensional primitives possessing, among other attributes, a position, shape, orientation, color, and transparency. Their combination restores the appearance of an object when observed from different viewpoints.

Idol offers several levels of detail, ranging from 32,768 to 1,048,576 Gaussians. Users can thus choose a lighter or more detailed representation according to the constraints of their project.

The examples presented by Speridlabs show that an object composed of 32,768 Gaussians occupies approximately 0.4 MB, compared to 3.2 MB for 262,144 Gaussians and 12.2 MB for the configuration exceeding one million.

The model exports its results in the standard `.ply` format, allowing them to be opened in tools compatible with Gaussian Splatting. However, it does not directly produce a polygonal mesh ready for use in all design software or for 3D printing. Conversion and potential corrections may be required for these uses. A compact model, but not as simple as it seems

One of the highlighted features concerns the size of the model. Idol comprises about 659 million parameters for its main generative components, divided between a 370.6-million-parameter diffusion Transformer and two decoders totaling about 288 million.

The complete system, however, uses several complementary models. DINOv3 is involved in visual analysis, FLUX.2's VAE participates in image encoding, and BiRefNet handles background removal.

Taking these dependencies into account, a full execution mobilizes approximately 1.8 billion parameters. The announced peak video memory consumption reaches 7.2 GB, theoretically allowing for usage on certain consumer graphics cards.

This architecture sets Idol apart from several commercial 3D generation solutions that rely mainly on remote platforms. With sufficient graphics memory and the required dependencies, the model can be installed and run locally.

Developers can also choose between two input modes: real, intended for photographs, and synthetic, designed for renders and artificially generated images. These modes adapt, among other things, the framing used during image preparation. Three seconds on a professional accelerator, longer on a PC

Speed constitutes another selling point for Speridlabs, but results vary significantly depending on the hardware.

In its performance benchmarks, the company announces a median time of 3.4 seconds to generate an object composed of 262,144 Gaussians on an NVIDIA B300 accelerator. The same process requires 6.7 seconds on an H100.

On consumer cards, processing times are longer. An RTX 5090 completes the operation in 14.1 seconds, compared to 19.1 seconds for an RTX 4090 and 46.2 seconds for an RTX 3080.

These results stem from tests conducted on eight images, with 50 generation steps and an identical configuration across the different cards. They apply to the level of detail of 262,144 Gaussians, and not systematically to the generation of one million Gaussians.

The presentation page also mentions processing times under two seconds in certain optimized configurations, but does not provide a comparable protocol to extend this performance across all graphics cards and all levels of detail. Encouraging results against TRELLIS and TripoSplat

To contextualize Idol's performance, Speridlabs published a comparison with three other 3D generation approaches: TripoSplat, TRELLIS, and LGM.

The evaluations cover object fidelity from the original viewpoint as well as the quality of newly generated views. Researchers use several metrics, including SSIM, LPIPS, IoU, DINOv2, and CLIP.

In these internal tests, Idol achieves a DINOv2 similarity score of 0.781 for new views, compared to 0.743 for TripoSplat, 0.754 for TRELLIS, and 0.575 for LGM. Its CLIP score reaches 0.885, ahead of the other three evaluated models.

On the LPIPS metric, which