Blender Idol

Marco Capelli in Rendering

Add-on for Blender · AI Gaussian splats

IDOL

Name: Blender IDOL
Brand: Marco Capelli
Price: 12.00 USD
Availability: InStock

Image to 3D

Powered by Idol, the brand-new single-image-to-3D model from Speridlabs. Drop in one picture and get a full 3D Gaussian splat, generated on your own NVIDIA GPU, right inside the viewport.

No cloud. No credits. Nothing to upload.

● Runs 100% locally● One-click install● ~2.5 min per object● 8 GB VRAM● Ready for Blender 5.3 native splats

−33%

LAUNCH PRICE

$18 $12

Until november 24

Meet Idol · the model behind it

One picture in. A full 3D splat out.

Brand-new: released by Speridlabs on October 28, 2026

Most Gaussian-splat add-ons display or edit splats that you capture elsewhere, from dozens of photos or a scan. This add-on uses Idol, a new open model that creates the splat for you from one single image. It even generates the sides the picture never showed.

0.66B parameters	< 8 GB GPU memory
1M Gaussians max	Open Apache-2.0 weights

State of the art at launch. In Speridlabs' published benchmark it beats TripoSplat, TRELLIS and LGM on 6 of 7 quality scores: SSIM, LPIPS, IoU, DINOv2, CLIP and sharpness. Outputs 32,768 to 1,048,576 Gaussians per object, written as a standard 3DGS .ply.

New-view quality	Idol	TRELLIS	TripoSplat	LGM
DINOv2 ↑	0.781	0.754	0.743	0.575
CLIP ↑	0.885	0.868	0.868	0.784

Mean scores on 1 × H100, from speridlabs.com/research/idol.

Model page: huggingface.co/speridlabs/idol · Project: speridlabs.com/research/idol
This add-on is an independent integration. It is not made or endorsed by Speridlabs.

How it works

Pick an image. Press Generate.

01 · ONE IMAGE → ONE SPLAT Pick, Generate, Done Choose an image file, or any image already in Blender: a render, a packed or generated image, or the Image Editor image. Press Generate 3D Splat. Background removal is automatic (BiRefNet); if the image already has transparency, that is used instead.	02 · REAL SPLATS, NOT PREVIEWS Up to 1 Million Gaussians Choose from 32K to 1M Gaussians. Each result is a real Blender object with a Geometry Nodes viewer and a material, and it renders in Eevee and Cycles. A matching camera is created from the predicted input view. Synthetic mode for renders and AI images, Real Photo for photographs.
03 · READY FOR BLENDER 5.3 Generate Here. Render Natively. Blender 5.3 (November 10, 2026) adds native splat import and rendering, but it only displays splats: you still need one to start with. Idol makes one from a single image, right inside Blender. Every result is saved as a standard 3DGS `.ply`, and on Blender 4.5–5.2 the add-on's own viewer and .ply import/export work today.	04 · SPLAT → TEXTURED MESH Need Polygons? Convert any splat into a UV-unwrapped mesh with a baked texture: 5K to 1M faces, 1K/2K/4K texture, floaters removed. Export it straight to `.glb`. Experimental: the side facing the camera converts well, the unseen back can have holes.

05 · Packed with extras

Everything you need, nothing to upload.

› AI Enhance Texture: Real-ESRGAN upscaling makes mesh textures noticeably crisper. Runs locally too.
› Live colour grading: exposure, temperature/tint, contrast, shadow lift, hue, saturation and gamma, live in the viewport.
› Sharpen & Clarity: inside each UV island, so no seams. One click restores the original bake.
› Point Cloud Viewport: about 46× faster per frame on a 1M-Gaussian splat in our test. Renders still use the full splat.

› Splat display controls: display ratio, splat size, minimum opacity and detail, live in the panel.
› .ply import & export: bring in splats from any 3DGS tool, or export yours with their transform.
› .glb export: textured meshes go out with your colour grading baked in.
› Saved Results manager: delete all results, or only the old ones, in one click from the Preferences.

Performance

How long does it take?

12–22 s

load model · once per session

~2:30

generate · 50 steps, 262K

~5 GB

peak GPU memory

Measured on an RTX 5060 Ti 16 GB. Fewer steps are faster. Keep the model loaded between generations, so you only pay the load once.

100% local

Install once, then offline.

› One-click install: its own private Python and PyTorch; nothing touches Blender's Python. CUDA version detected automatically.
› Download the models: about 6.6 GB, with progress bar, speed and resume.
› Free Hugging Face account: Meta's DINOv3 is gated. Accept its terms once and paste a Read token; the add-on links to the page.
› Disk: about 18 GB total after install.

› Blender never freezes: generation runs in a separate background process, with live progress, a step strip and Cancel.
› Multi-GPU aware: pick which NVIDIA GPU to use. Driver and VRAM are checked up front.
› Seed control: same seed, same object. Or randomise.
› Advanced controls: steps, guidance, time shift and crop margin.
› Private by design: your images never leave your computer.

Requirements

Recommended system

Component	Requirement
GPU	NVIDIA (CUDA), RTX 20 series or newer, 8 GB VRAM
RAM	16 GB
OS	Windows 10/11 x64
Blender	4.5 – 5.2
Disk	~20 GB free

Credits & licences

The add-on code is GPL-3.0-or-later. It uses Idol by Speridlabs (Apache-2.0). It is Built with DINOv3 (Meta, DINOv3 License) and uses the FLUX.2 autoencoder by Black Forest Labs (Apache-2.0 release). Full notices are in THIRD_PARTY_NOTICES.md. Model weights are downloaded from their official pages and never redistributed.