Mira-Scene, from VAST-AI-Research, reconstructs an editable 3D scene from a single image. Its approach is compositional: object geometry is generated separately and then placed into the scene, and the paper replaces sparse pose regression for that placement with dense, bounded correspondence recovery. The core is the Canonical Coordinate Map (CCM), a pixel-aligned field that maps each visible object pixel to a surface coordinate in the object's bounded canonical space. Paired with a scene-space Point Cloud Map from monocular geometry estimation, it yields canonical-to-scene correspondences from which object transformations are recovered by robust geometric alignment, and the paper describes a multimodal diffusion transformer that jointly generates object geometry and CCMs. The stage-wise inference pipeline covers segmentation, depth, CCM, mesh, floor, scene and environment, with resumable execution, multiple depth backends (MoGe, MoGe-2, PPD), mesh backends (SAM3D, TRELLIS.2) and case-level multi-GPU sharding; it ends with an equirectangular environment map. The README lists a comparison table against SAM3D: 3D-IoU 0.727 versus 0.520 and 2D-IoU 0.783 versus 0.672, with Mira-Scene trained on 80k 3D data versus SAM3D's stated million-plus, though SAM3D is ahead on EMD (0.163 versus 0.169). The paper abstract reports relative gains of 39.8% in 3D-IoU and 16.5% in 2D-IoU over SAM3D. The repository includes the inference pipeline, Mira-CCM, UniDataset, evaluation and visualization tools, example cases, and training configs for a pretraining stage and a finetuning stage. Stage 1 and Stage 2 checkpoints are published on Hugging Face under Yang-Tian/Mira-Scene, and the finetuning data is released as the Mira-Scene Dataset, though the large Objaverse rendering data used for pretraining is not included. The code is MIT-licensed; third-party components, pretrained models and datasets keep their own licenses and terms. Inference stages need separate Python, PyTorch and CUDA environments.