speridlabs Releases "Idol," a Model That Generates Up to About 1.05 Million 3DGS Gaussians From a Single Image

What happened

speridlabs has released "Idol," a model that generates 3D Gaussian Splatting (3DGS) from a single image of an object, on Hugging Face. The input can be a rendered image, an AI-generated image, or a photograph, and the output is a common 3DGS .ply file. According to the publisher, it generates up to 1,048,576 Gaussians, including sides that are not visible in the input.

The pipeline consists of background removal, object cropping, and Gaussian generation. The publisher says it runs on consumer GPUs, uses 7.2GB of VRAM, and can run in under 2 seconds under certain conditions.

The code is in the "idol-sdk" repository on GitHub, and the weights (about 3GB) are downloaded automatically from Hugging Face on first run. The page shows 20,689 downloads in the last month, 6 derived models, 4 quantized versions, and 14 Spaces that use this model.

How it works

The model is made up of several parts. The core components are a DiT (370.6M parameters) that handles denoising, an Octree decoder (50.8M), and a Gaussian decoder (237.3M).

Input processing uses BiRefNet (220.7M) for background removal, DINOv3 ViT-H+/16 (840.6M) for image feature extraction, and the FLUX.2 VAE (84.1M). The whole model has about 1.80B parameters.

Inference uses diffusion model sampling, with a default of 50 steps. Reducing the steps makes it faster, but fine detail is reportedly lost. cfg_strength (default 3.0) adjusts fidelity to the input image, and rescale_t (default 3.0) adjusts whether more time is spent on overall shape or on fine detail.

On GPU, speed-up settings are enabled by default. The denoiser is compiled with torch.compile, guided and unguided passes are combined into one batch, and attention uses cuDNN and TF32. The publisher says this makes it about 1.7x faster while preserving output quality. The first image is slower because of compilation.

There are two input modes. synthetic (default) is for renders and generated images, and crops so that the object occupies 1/1.2 of the frame. real is for photographs, and leaves wider margins so that the object occupies 1/1.7.

By the numbers

Why it matters

For game, XR, and prototype production, this adds an option for creating 3DGS assets from a single image in a short time. Because the output is a standard .ply, it is easy to pass to existing splat viewers and conversion tools.

Because the weights and code are public, it can be tried on a local GPU. Even for teams that do not want to send data to a cloud API, evaluation in their own environment becomes easier.

However, the generated back and side surfaces are estimated from the image. For uses that require accurately recording the real shape and color of an object (cultural heritage documentation, inspection, measurement, etc.), it should be kept in mind that this differs in nature from reconstruction from multiple captured images.

How to try it

1. With your Hugging Face account, accept the terms of use for facebook/dinov3-vith16plus-pretrain-lvd1689m and black-forest-labs/FLUX.2-dev (only the VAE is downloaded), then log in with hf auth login.
2. Clone idol-sdk from GitHub and install it with pip install -e .
3. For photos, run python -m idol.sample --image photo.jpg --mode real. For renders or generated images, specify a folder and use --mode synthetic.
4. Open the output .ply in any Gaussian splat viewer to check it. You can also check the background removal and crop results in <name>_input.png.
5. To reproduce the same result, fix --seed. The number of Gaussians can be changed with --num_gaussians.

Caveats and open questions

• Benchmarks and speeds are the publisher's own measurements. Third-party verification cannot be confirmed from the source material.
• The conditions for the "under 2 seconds" figure are not stated in the source material. In the table provided, the fastest result at 262,144 Gaussians and 50 steps is 3.4 seconds on the B300. Results vary greatly with environment and settings.
• Reported times per GPU include 46.2 seconds on the RTX 3080 and 54.6 seconds on the L4. Even with a peak VRAM of 7.2GB, practical speed depends on the GPU.
• The benchmark metrics are PSNR, SSIM, and others for the input view, and DINOv2, CLIP, and sharpness distance for novel views. The visual quality of the generated output cannot be judged from these numbers alone.
• Use requires agreeing to the terms of two gated models, DINOv3 and FLUX.2-dev. The license terms for these and for Idol itself cannot be confirmed from the source material. Check the terms on each page before actual use.
• An NVIDIA GPU (CUDA) is required. Operation on AMD GPUs or Macs is not mentioned in the source material.
• Links to a paper or technical report cannot be confirmed from the source material. Details of the training data and training method have not been published.

Glossary

3D Gaussian Splatting (3DGS)

A 3D representation that expresses a scene with many semi-transparent 3D Gaussians (ellipsoids) and renders it quickly. It is stored in files such as PLY.

PLY

A file format widely used to store point clouds, meshes, and 3DGS.

DiT

Short for Diffusion Transformer. A denoising (diffusion) model that uses a Transformer.

Classifier-free guidance (CFG)

A technique for adjusting how strictly the generated result follows the condition (here, the input image).

LPIPS / PSNR / SSIM

Metrics for measuring image fidelity. Higher PSNR and SSIM, and lower LPIPS, are considered closer to the original image.

VRAM

Memory on the GPU. It serves as a guide to which GPUs can run the model.

Sources

Primary sources (publisher, paper, repository)

• speridlabs/idol · Hugging FaceHugging Face（speridlabs）（huggingface.co）・en

The X post that surfaced this story

• View @Z3nlyte's post on Xx.com

This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original