VAST open-sources Mira-Scene to turn single images into editable 3D worlds

Mira-Scene is an open-source pipeline that reconstructs fully editable 3D scenes from a single 2D photograph.

Most existing generative systems try to assemble 3D environments from scratch, but individual objects almost always drift in scale, orientation, and placement relative to the camera. Mira-Scene fixes this by enforcing dense, pixel-aligned correspondences between the 2D input and the reconstructed 3D world.

What it does

Feed it an ordinary image, and Mira-Scene automatically segments foreground instances, predicts depth, builds clean 3D meshes for each object, and places them into an assembled scene with a rendered environment map.

Because each item exists as an independent mesh, you are not stuck with an uneditable static point cloud. You can export the whole setup straight into Blender, tweak individual furniture pieces, add lighting, or export it to physics simulators for training robotics agents.

Under the hood

Rather than relying on closed agentic loops, Mira-Scene uses a multi-stage modular architecture. It starts with instance segmentation and depth estimation, followed by canonical coordinate and voxel prediction powered by Mira-CCM.

The pipeline then generates object-level geometries and performs automatic scene assembly, recovering precise 3D bounding boxes and camera poses to match the original pixels.

Even more impressive are the benchmark numbers. Compared against closed-source competitors like SAM3D—which relies on over a million assets and heavy human annotations—Mira-Scene was trained on an automated pipeline with just 80k data samples. Yet it beats SAM3D across the board, jumping from 0.520 to 0.727 in 3D-IoU and from 0.672 to 0.783 in 2D-IoU.

How to get started

The code is released under the MIT license on GitHub, and the entire download footprint is roughly 6.1GB, making it lightweight enough to run on standard mid-range consumer GPUs.

VAST AI Research has published the full pipeline, checkpoint models for both pretraining and finetuning, and the finetuning datasets on Hugging Face. The pipeline supports resumable runs, interchangeable depth and mesh backends, and multi-GPU distributed batching.

Why it matters

For 3D artists, game developers, and robotics researchers, Mira-Scene offers an immediate practical boost. Turning a single snapshot into separated, simulation-ready 3D assets previously required hours of tedious photogrammetry cleanup or manual modeling.

It also proves that tight geometric alignment beats brute-force training scale. Beating closed-source million-scale models with just 80,000 automated samples makes Mira-Scene one of the most promising open foundational tools for generative spatial computing.