Idol turns single 2D images into 3D Gaussians on consumer GPUs

Idol is a new open-source 3D generation tool from Sperid Labs that transforms any single 2D image into a standard 3D Gaussian splat.

Whether you feed it a real camera photo, a 3D render, or an Iris-3B AI generation, Idol strips the background, infers what the unseen backside looks like, and outputs a complete .ply file packed with up to a million 3D Gaussians. Even better, peak VRAM is just 7.2 GB, meaning it runs entirely on local consumer hardware.

What it does

Instead of generating heavy, noisy polygon meshes that often melt on the hidden sides, Idol scatters calibrated Gaussian points across space to represent volume and surface texture.

Users can dial the output resolution between 32,768 and 1,048,576 Gaussians depending on their needs. A low-density 32k splat gives an instant spatial preview, while a full million-point generation recovers sharp details across the object.

In official benchmarks run on an H100 GPU against TRELLIS, TripoSplat, and LGM, Idol came out on top across novel unseen viewpoints. It scored highest on DINOv2 and CLIP similarity tests while posting the lowest sharpness distance at 0.81. While LGM showed a higher raw PSNR, its results suffer from noticeable blurriness, whereas Idol maintains crisp geometry on both front and back.

How it runs and system requirements

Idol stays lean by relying on a modular pipeline rather than a giant monolithic transformer. Its core generative denoiser and decoders total around 0.66 billion parameters. Even with BiRefNet for matting, Meta's DINOv3 for visual features, and FLUX.2's VAE loaded, total weights stay around 4 GB.

Memory management is remarkably tight. With built-in torch.compile and TF32 sampling enabled, peak VRAM caps at 7.2 GB. Generation is fast: an enterprise B300 chip handles a 262k splat in 3.4 seconds, with the fastest runs touching roughly 2 seconds. On consumer desks, an RTX 5090 takes 14.1 seconds, an RTX 4090 clocks in at 19.1 seconds, and an older 10 GB RTX 3080 finishes in roughly 42 to 46 seconds.

Both code and weights are public right now. You can clone the SDK from GitHub (speridlabs/idol-sdk) and pull weights from Hugging Face. Setup requires Python 3.10+, PyTorch 2.5+, and a CUDA-capable GPU.

One small catch: Idol reads inputs using two gated Hugging Face models—DINOv3 ViT-H+/16 and FLUX.2-dev (VAE only). You will need to accept their terms on Hugging Face and log in via your terminal before the first run.

Why it matters

For indie game developers, digital artists, and creative technologists, Idol removes the friction of spinning up cloud clusters just to prototype a 3D prop. Being able to test concepts on an everyday 8 GB gaming card turns 3D asset drafting into an instant workflow.

We think Sperid Labs made a very smart architectural bet here. Instead of bloating the model with billions of generic weights, they offloaded semantic understanding to DINOv3 and latent reconstruction to FLUX, keeping their 3D denoiser small, agile, and local-friendly.

Keep in mind that 3D Gaussians are not traditional rigged meshes; taking these assets into a standard game engine physics loop will still require mesh conversion. But for fast spatial mockups, virtual product visualization, and interactive scene building, Idol is easily one of the most accessible open-source tools released this season.