📄 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.

🔥 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.

The stage-wise pipeline supports resumable execution, multiple depth and mesh backends, and case-level multi-GPU sharding. See the inference guide (中文).

See the visualization guide (中文).

See the evaluation guide (中文).

Please refer to the evaluation guide for details.

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.

@article{sun2026mira,
  title={Mira-Scene: Pixel-Aligned Layouts for Generative 3D Scene Reconstruction},
  author={Sun Yang-Tian and Liu Tianjia and Huang Zehuan and Huang Yi-Hua and Lyu Xiaoyang and Yang Ziyi and Zou Zi-Xin and Guo Yuan-Chen and Cao Yan-Pei and Qi Xiaojuan},
  journal={arXiv preprint arXiv:2609.23796},
  year={2026}
}

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.