idol 3D Gaussian Splatting Generation Model: File List

Sample outputs are available on the model card.
At a Glance
| Item | Value |
|---|---|
| Repository | speridlabs/idol |
| Family guide | idol guide (2 articles) |
| Published | 2026-10-29 |
| License | apache-2.0 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
speridlabs has released “Idol", a model that generates 3D Gaussian Splatting (3DGS) data from a single image.
When given a single object image (such as a photograph, CG rendering, or AI-generated image), this model automatically performs background removal and composition adjustment, outputting up to 1,048,576 3D Gaussians including the hidden backside not visible in the image. The generation results are exported as .ply files in the standard 3DGS format.
Specifications
Model architecture, specifications, and recommended settings described in the public documentation are as follows:
- Total Parameters: Approx. 1.80B
- Component Configuration:
- DiT (denoiser): 370.6M (FP16 transformer blocks, FP32 embedding and normalization layers) – Octree decoder: 50.8M (FP16 transformer blocks, FP32 embedding and normalization layers) – Gaussian decoder: 237.3M (FP16 transformer blocks, FP32 embedding and normalization layers) – BiRefNet (background removal): 220.7M (FP16) – DINOv3 ViT-H+/16 (image encoder): 840.6M (bf16) – FLUX.2 VAE: 84.1M (bf16, cast from FP32 checkpoint upon loading)
- Output Specifications:
.plyfile in standard 3DGS format, along with_input.pngfor the input image (after background removal and cropping) and_camera.jsonfor estimated camera parameters - Number of Output Gaussians (
--num_gaussians): Default is 262,144 (can be specified in multiples of 32 from 32,768 to 1,048,576) - Recommended Sampling Steps (
--steps): 50 steps (reducing steps speeds up generation but loses details) - Guidance Scale (
--cfg_strength): Default 3.0 (Classifier-free guidance, controlling how faithfully the model follows the image) - Time Shift (
--rescale_t): Default 3.0 (higher values allocate more steps to the overall shape, lower values allocate them to fine details) - Input Mode (
--mode):synthetic(default, for CG renderings and generated images where the object occupies 1/1.2 of the frame) andreal(for photographs where the object occupies 1/1.7 of the frame) - Encoder Precision (
--dinov3_dtype): Default isbf16(specifyingfp32runs the DINOv3 encoder in full precision) - Optimization Options: Fast sampling is enabled by default (denoiser compilation via
torch.compile, batching of guided and unguided passes, and attention processing using cuDNN and TF32 matrix multiplication, resulting in about 1.7x faster speed while maintaining output quality) - License Terms: The overall repository is licensed under the Apache 2.0 license. However, the image encoders used for input processing (
facebook/dinov3-vith16plus-pretrain-lvd1689mandblack-forest-labs/FLUX.2-dev) are restricted models that require agreeing to individual terms of use on Hugging Face
Performance
In benchmark evaluations published by the creators, scores were measured for Idol, TripoSplat, TRELLIS, and LGM across input views (PSNR, SSIM, LPIPS, IoU) and 7 novel views (DINOv2, CLIP, Sharpness distance). Measurements were conducted using 1 x H100 per model.
→ Scroll horizontally to see all columns
| Model | Setting | PSNR ↑ | SSIM ↑ | LPIPS ↓ | IoU ↑ | DINOv2 ↑ | CLIP ↑ | Sharpness distance ↓ | Request time ↓ |
|---|---|---|---|---|---|---|---|---|---|
| Idol | 50 steps, 262,144 Gaussians | 12.25 | 0.726 | 0.213 | 0.862 | 0.781 | 0.885 | 0.81 | 6.8 s |
| TripoSplat | 50 steps, 262,144 Gaussians | 11.76 | 0.712 | 0.225 | 0.860 | 0.743 | 0.868 | 1.03 | 9.9 s |
| TRELLIS | Paper setting, 50 + 50 steps | 11.89 | 0.716 | 0.235 | 0.845 | 0.754 | 0.868 | 0.97 | 7.1 s |
| LGM | Default, 65,536 Gaussians | 12.46* | 0.722 | 0.237 | 0.834 | 0.575 | 0.784 | 1.22 | 1.1 s |
According to the published figures, Idol achieves the best scores among the compared models in many geometric and perceptual metrics, such as SSIM (0.726), LPIPS (0.213), IoU (0.862), and novel-view DINOv2 (0.781), CLIP (0.885), and Sharpness distance (0.81). Meanwhile, its PSNR stays at 12.25 compared to LGM’s 12.46, and its request time is 6.8 seconds (faster than TripoSplat’s 9.9 seconds, but slower than LGM’s 1.1 seconds and TRELLIS’s 7.1 seconds). According to notes from the authors, LGM yields the highest PSNR due to the metric’s characteristic of averaging squared pixel errors, despite producing visually blurry results.
Additionally, the creators provided the following measurement results for request times (median excluding initial compilation for 8 images) when generating 50 steps and 262,144 Gaussians across various GPUs:
| GPU | VRAM | Request time |
|---|---|---|
| B300 | 269 GB | 3.4 s |
| B200 | 179 GB | 3.6 s |
| H200 | 140 GB | 6.6 s |
| H100 | 80 GB | 6.7 s |
| RTX 5090 | 32 GB | 14.1 s |
| A100 | 80 GB | 16.3 s |
| L40S | 45 GB | 16.6 s |
| RTX 4090 | 24 GB | 19.1 s |
| RTX 3080 | 10 GB | 46.2 s |
| L4 | 22 GB | 54.6 s |
Strengths and Use Cases
Idol excels at rapidly reconstructing 3D structures spanning the entire surroundings of an object from a single still image. Main intended use cases and features include:
- 3D Asset Generation from a Single Image: From just one image, it infers shaded surfaces and backsides not captured in the input, generating three-dimensional and robust 3D Gaussian data. It can be utilized for game production, VR/AR asset prototyping, and shape evaluation for 3D printing.
- Broad Input Support for Photos, CG, and Generated Images: High-precision 3DGS data can be generated not only from live-action photographs but also from 3D rendered images and 2D illustrations or concept art created by generative AI.
- Automatic Composition Optimization Based on Input: Background removal is performed inside the model via BiRefNet, and cropping margins for the subject are automatically adjusted according to two modes (
syntheticandreal) matched to the input source. This eliminates the effort of manual background cropping or composition normalization on the user side. - Detail Level Adjustment According to Use Case: The number of generated Gaussians can be specified in multiples of 32 from 32,768 up to 1,048,576. The output volume can be controlled depending on the purpose, ranging from low-density data for environments requiring lightweight display to high-quality assets requiring precise detail reproduction.
- Output in Standard Format and Interoperability: Because generation results are exported in the industry-standard 3DGS
.plyformat, they can be directly imported into existing general Gaussian Splatting viewers, renderers, and external tools. Furthermore, background-removed images (_input.png) and estimated camera information (_camera.json) are output simultaneously, facilitating integration with downstream pipelines.
Distributed Files
Weight files published in speridlabs/idol, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
birefnet.safetensors |
444MB |
decoder.safetensors |
203MB |
decoder_gs.safetensors |
949MB |
denoiser.safetensors |
1.48GB |
How to Get It
Idol’s code and model are available on GitHub and Hugging Face. Weight files are distributed in safetensors format and are retrieved automatically via the SDK.
As an important prerequisite for using this model, you must agree to the terms of use for two external encoders. While the Idol codebase itself is released without restriction under the Apache 2.0 license, the following two models used for input processing require agreeing to terms of use (gated access approval) on Hugging Face:
facebook/dinov3-vith16plus-pretrain-lvd1689mblack-forest-labs/FLUX.2-dev(only the VAE portion is used)
Make sure to agree to each model’s terms of use on your Hugging Face account in advance, then run hf auth login in your terminal to complete authentication, or set the HF_TOKEN environment variable. If executed without access rights, model downloading will not start and an error message indicating missing permissions will be displayed.
The environment setup procedure is as follows. Execution requires Python 3.10+, PyTorch 2.5+, and a CUDA-capable NVIDIA GPU.
git clone https://github.com/speridlabs/idol-sdk.git
cd idol-sdk
pip install -e.
There is no need to manually download weight files separately. Upon the first run, Idol’s weights and the aforementioned image encoders will be automatically downloaded to the Hugging Face cache directory.
To generate from the CLI, run the following command. Specify real (for photos) or synthetic (for CG/generated images) for --mode depending on the nature of the input image:
# Generating from a photograph
python -m idol.sample --image photo.jpg --mode real
# Generating from a rendering image directory and specifying the output destination
python -m idol.sample --image renders/ --mode synthetic --out outputs
It is also possible to call the pipeline directly from a Python script. Load and generate using idol.IdolPipeline as follows:
from idol import IdolPipeline
pipeline = IdolPipeline.from_pretrained("speridlabs/idol")
output = pipeline("photo.jpg", mode="real", seed=42, num_gaussians=1_048_576)
output.gaussian.save_ply("photo.ply")
During generation, optional flags such as seed (--seed), sampling steps (--steps, default 50), guidance strength (--cfg_strength, default 3.0), time shift (--rescale_t, default 3.0), and the number of Gaussians to generate (--num_gaussians, default 262144) can be adjusted as needed.
Can You Run It Locally?
The publisher distributes this model as safetensors.
License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Related Articles
What to Read Next
- Explore the same model family → idol family overview (2 articles)
- Formats this model is available in → Safetensors format guide and models

