Speridlabs' Idol Turns a Single Photo Into a 3D Model on Consumer GPUs
Speridlabs released Idol , a single-image to 3D Gaussian splat model under Apache 2.0. Up to 1,048,576 Gaussians per object, 7.2 GB VRAM peak, under 2 seconds on datacenter GPUs. Diffusion transformer (370M DiT) with DINOv3 and FLUX.2 VAE image conditioning, BiRefNet background removal.
Takeaways
- Speridlabs released Idol , a single-image to 3D Gaussian splat model under Apache 2.0.
- Up to 1,048,576 Gaussians per object, 7.2 GB VRAM peak, under 2 seconds on datacenter GPUs.
- Diffusion transformer (370M DiT) with DINOv3 and FLUX.2 VAE image conditioning, BiRefNet background removal.
- Beats TripoSplat, TRELLIS and LGM on SSIM, LPIPS, IoU and novel-view DINOv2/CLIP scores.
- Runs on RTX 3080 (46 s) up to B300 (3.4 s); try the HF Space or fal API .
- Outputs 3DGS
.ply, not editable mesh; built for single objects, not scenes or riggable characters.
Speridlabs has released Idol , a single-image 3D generator that runs locally with a reported 7.2 GB VRAM footprint. Given a photograph, render, or AI-generated image of an object, it produces a standard 3D Gaussian splat file with predicted geometry and appearance for views outside the original camera angle. The SDK code carries an Apache-2.0 license.
A million Gaussians from one view
A 3D Gaussian splat represents an object as a collection of colored, semi-transparent ellipsoids. A compatible viewer projects those primitives quickly enough for interactive camera movement, without first converting the result into a polygon mesh.
Idol removes the background, frames the object, and generates between 32,768 and 1,048,576 Gaussians. The default is 262,144. It writes a 3DGS .ply file for splat viewers and companion JSON containing the predicted camera for the source image.
Because the model receives one view, it predicts surfaces hidden from the camera instead of reconstructing them from evidence. The resulting back and side views may look plausible while differing from the original object.
7.2 GB, with timing caveats
Idol’s reported 7.2 GB VRAM use puts local inference within reach of many gaming GPUs. The lowest-memory card in Speridlabs’ published tests is a 10 GB RTX 3080, so the available data does not establish compatibility with 8 GB cards.
Speridlabs measured median request times with 262,144 Gaussians and 50 Euler sampling steps across the same implementation:
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
