Idol on LiteRT

Idol on LiteRT

Turns one photo of an object into a 3D Gaussian splat on an Android phone, using LiteRT versions of the Idol model and a sample app, for mobile and 3D developers.

Made by litert-community on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository packages Idol for LiteRT, so one photo of an object becomes a 3D Gaussian splat on a phone. You can turn the result around and see sides the photo never showed. It holds the denoiser with the idol-turbo LoRA, the decoders, a VAE encoder and a background remover. The DINOv3 image encoder lives in a separate gated repository.

Who it suits: Android developers who want image-to-3D running on the device will find it fits, as will 3D and machine learning engineers who study converted LiteRT graphs. Expect minutes, not seconds, per photo.

Industry | Who uses it | How they use it | Example
Mobile apps | Android developers | Run photo-to-splat on the phone with the sample app | Building a scan feature that works offline
3D content | Creators | Make a turnable 3D view of one object from a single photo | A quick 3D look at a plant or a chair
Research | ML engineers | Study how the model was converted into LiteRT graphs | Comparing the two forms on speed and accuracy
Hardware testing | Device teams | Measure heat and speed on real phones | Timing a run from a cool phone

If you need a splat in seconds, you can skip it, since the card's own timings run to minutes.

Where to use it:

• An Android phone, using the sample app in the android folder
• LiteRT, through its CompiledModel interface
• A Python runner for the same two forms
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Photo to 3D | Android developers | A .ply file holding the splat
Speed testing | Device teams | Timings for each stage on a given phone
Form comparison | Researchers | Speed and closeness scores for two forms

How to set it up: You need an Android phone, the sample app, and access to the separate DINOv3 repository. The card measured a Galaxy S26 and a Pixel 11 Pro with LiteRT 2.3.0.

• An Android phone, such as the ones the card measured
• LiteRT 2.3.0, the version the card used
• Access to the gated DINOv3 repository, which has its own licence
• Python with pip, to download the files

1. Install the Hugging Face hub package.
2. Download the repository's files.
3. Get the DINOv3 encoder from its own gated repository.
4. Build and run the sample app from the android folder.

pip install -U huggingface_hub
hf download litert-community/Idol

How to use it: Give the sample app a photo of one object. It removes the background, encodes the image, runs the denoiser and decoders, and writes a .ply file. The text shown here gives no build commands, so follow the model page for those.

Tool | What it does
Default form (repo root) | Makes 65,536 Gaussians and is the fastest form the card measured
262k folder form | Makes 262,144 Gaussians and matches the author's default output more closely
android folder sample app | Turns a photo into a splat on the phone
Python runner | Runs both forms, pointed at the root or the 262k folder

Example: You open the sample app on a Galaxy S26 and give it a photo of a plant in a flowerpot. The card timed this at 158 seconds on a cool phone. The result is a splat of 65,536 Gaussians that you can rotate to inspect.

For newcomers and experts: A Gaussian splat is a 3D scene made of many soft coloured blobs, which lets you look at an object from other angles. A good first step is to read the card's two forms and pick the default one.

The default form shares one graph across the denoiser's middle blocks and across the decoder's middle graphs, copying weights in before each call. The card lists dead ends too. Eight-bit weights failed on the phone GPUs, and the denoiser could not be built for the S26's NPU. A warm phone slows runs a lot, and the default form agrees less closely with the author's output than the larger form does.

You give the app one photo of an object
BiRefNet removes the background on the CPU
The encoders and the denoiser run on the GPU
The decoders write a 3D splat as a .ply file

Title: Time per photo, first photo after start, in seconds
Default form on Galaxy S26: 158 seconds
First release form on Galaxy S26: 267 seconds
Default form on Pixel 11 Pro: 343 seconds

Good to know:

• License: apache-2.0, as listed on Hugging Face
• The DINOv3 encoder sits in a separate gated repository under its own licence
• Cost to run: none named
• Data it reaches: the photo you give the app, and the source names no web or account access
• Not for: quick results, since the card's timings run from about 158 to 449 seconds
• Heat matters, because a warm phone took longer in the card's tests

On Hugging Face, 13 November 2026

• Task: image to 3d
• Library: litert
• Downloads in the last 30 days: 1,066
• Likes: 5
• License: apache-2.0
• Based on: speridlabs/idol
• Tags: litert, tflite, android, on-device, gaussian-splatting, 3d, image-to-3d

Links

• Model: https://huggingface.co/litert-community/Idol
• Files: https://huggingface.co/litert-community/Idol/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download litert-community/Idol

Then follow the usage in its model card. Nothing in this kit runs by itself.