VAST-AI-Research / Mira-Scene

# Mira-Scene: Pixel-Aligned Layouts for Generative 3D Scene Reconstruction
📄 Paper | 🌐 Project Page
Mira-Scene reconstructs an editable 3D scene from a single image. The pipeline
combines interactive instance segmentation, depth estimation, canonical
coordinate and voxel prediction, object mesh generation, scene assembly, and
environment-map generation.
Agentic systems such as GPT-6 Astra can assemble a scene from scratch, but the
generated objects drift in geometry, scale, and layout relative to the input
image. Mira-Scene instead recovers dense, pixel-aligned correspondences and
places each object so the reconstructed scene stays consistent with the source.
## 📰 News
> **🔥 Mira-Scene is fully open-source**
>
> The Mira-Scene inference pipeline, Mira-CCM, UniDataset, evaluation and
> visualization tools, example cases, and both the **Stage 1 pretrained
> checkpoint** and **Stage 2 final checkpoint** are publicly available in this
> repository and through Hugging Face.
>
> **✅ Finetuning data is now available**
>
> The released Scene data and outpainted Objaverse data are available from the
> Mira-Scene Dataset repository.
> The large Objaverse rendering data used for pretraining is not included;
> it can be generated conveniently from the source assets.
## Inference
The stage-wise pipeline supports resumable execution, multiple depth and mesh
backends, and case-level multi-GPU sharding. See the inference
guide (中文).
## Visualization
See the visualization guide
(中文).
## Evaluation
See the evaluation guide
(中文).
| Method | Data Availability | 3D Data Scale | Data Preparation | CD↓ | FS@0.1↑ | EMD↓ | 3D-IoU↑ | ICP-Rot↓ | 2D-IoU↑ | ADD-S↓ |
| --- | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| SAM3D | Closed-source | Million+ | Extensive manual layout annotation | 0.027 | 0.817 | **0.163** | 0.520 | 7.566 | 0.672 | 0.078 |
| Ours | Open-source | 80k | Fully automatic pipeline | **0.021** | **0.843** | 0.169 | **0.727** | **5.616** | **0.783** | **0.031** |
Please refer to the evaluation guide
for details.
## Out-of-the-box pipeline
Mira-Scene runs as a ready-to-use pipeline: from a single image it segments
instances, estimates depth, predicts canonical coordinates, generates object
meshes, assembles the scene, and renders an environment map — no per-scene
layout annotation or manual placement.
## Training
Training is organized into two stages:
1. Pretraining with `example_train/configs/pretrain_l1.yaml`;
2. Finetuning with `example_train/configs/finetune.yaml`.
A pretrained Mira-CCM checkpoint is available from
Hugging Face and can be used
directly for the second-stage finetuning.
See the training guide for installation,
configuration, checkpoint download, and launch instructions.
## Citation
 [code] 
## License
The Mira-Scene code in this repository is released under the MIT License.
Third-party components, pretrained models, and datasets retain their respective
licenses and terms of use.