InSpace: Structure-Aware 3D Indoor Scene Generation from a Single 360° Image

ECCV 2026 · Malmö, Sweden

Gwanhyeong Koo^{1,2 *}, Hyunsu Kim², Youngji Kim², Taejae Lee²,
Siwoo Lim¹, Sunjae Yoon³, Suyong Yeon^{2 †}, Chang D. Yoo^{1 †}

¹ KAIST    ² NAVER LABS    ³ Chung-Ang University
_{^* Work done during an internship at NAVER LABS.    ^† Co-corresponding authors.}

InSpace generates a complete, asset-aware 3D indoor scene from a single 360° (equirectangular) panorama, producing a full-room mesh along with individual, separable, textured furniture meshes. It is built on the TRELLIS.2 O-Voxel representation and extends it with a panorama-native, structure-aware generation pipeline: view-selective cross-attention driven by the camera center, layout-guided structure inversion from monocular depth, a 3D bounding-box estimator, and asset-aware shape and texture generation with global-local hybrid attention.

✨ Highlights

• Single 360° image to full 3D room. No multi-view capture, no per-scene optimization.
• Asset-aware output. The scene is decomposed into a layout (floor and walls) and individual objects, each exported as its own mesh, not a single fused blob.
• Structure-aware conditioning. A 360° panorama is unwrapped into 6 cubemap faces (FOV 120°), and each voxel attends only to the faces visible from its 3D position, via a camera-center conditioned view-selective cross-attention.
• Layout-Guided Structure Inversion. A monocular-depth (Depth-Anything-2) point cloud, the Partial Scene Geometry (PSG), seeds coarse generation via SDEdit-style noise inversion, improving room-scale fidelity.

🗓️ Roadmap

• ✅ Pretrained checkpoints on Hugging Face (InSpace)
• ✅ ERP-FRONT-30K dataset on Hugging Face (ERP-FRONT-30K)
• ✅ Interactive Gradio demos (ERP-FRONT / Structured3D / ReplicaPano / custom)
• ✅ Inference & evaluation pipeline
• ✅ Training scripts (scripts/train/)
• ⬜ Dataset preprocessing pipeline (data_toolkit/erp/)

🗺️ Pipeline

Single 360° ERP panorama
        │  depth estimation + ERP back-projection
        ▼
Stage 1 · Partial Scene Geometry (PSG)      spatial prior + calibrated camera center
        │  unwrap to 6 cubemap faces (FOV 120°), view-selective cross-attention
        ▼
Stage 2 · Coarse Scene Geometry (flow matching)   ->   voxel scene [1,64,64,64]
        │  3D bounding-box detector   ->   per-asset oriented boxes (OBBs)
        ▼
Stage 3 · Detailed Layout & Asset generation (global-local hybrid attention)
        ▼
textured scene mesh:  scene.glb  +  layout.glb  +  assets/{i}.glb

The four trainable components correspond to the four released checkpoints (see Pretrained Weights).

🛠️ Installation

InSpace uses the same environment as TRELLIS.2 (PyTorch 2.6 / CUDA 12.4 recommended).

Prerequisites

• OS: Linux. GPU: NVIDIA with at least 24 GB VRAM (tested on A100 / H100).
• CUDA Toolkit 12.4, Conda, Python 3.8 or newer.

Steps

git clone <this-repo-url> --recursive
cd InSpace

# Create the `inspace` conda env and install all dependencies
. ./setup.sh --new-env --basic --flash-attn --nvdiffrast --nvdiffrec --cumesh --o-voxel --flexgemm

Run . ./setup.sh --help for the full list of flags. For GPUs without flash-attn support (e.g. V100), install xformers and set ATTN_BACKEND=xformers.

📦 Pretrained Weights

Weights are not committed to this repo (size). Download them from Hugging Face into ckpts/ (the repo mirrors the local layout, so --local-dir . places them correctly):

hf download GwanHyeong/InSpace --include "ckpts/*" --local-dir .

InSpace uses four checkpoints plus the base TRELLIS.2 weights:

Folder (ckpts/…)	Component	Role
erp_ss_flow_img_dit_L_16l8_bf16_spatial/	Coarse geometry	Stage 2 coarse scene structure (sparse structure)
bbox_centerpoint/	3D BBox	Per-asset oriented bounding-box estimator
erp_slat_flow_img2shape_asset_aware_bf16/	Asset shape	Stage 3 asset-aware shape generation
erp_slat_flow_imgshape2tex_asset_aware_bf16/	Asset texture	Stage 3 asset-aware texture generation

The base TRELLIS.2 / TRELLIS-image-large weights are pulled automatically from Hugging Face (microsoft/TRELLIS.2-4B, microsoft/TRELLIS-image-large) as referenced in configs/gen/.

The processed dataset is released on Hugging Face as ERP-FRONT-30K; see datasets/README.md.

hf download GwanHyeong/ERP-FRONT-30K --repo-type dataset --local-dir datasets/

🚀 Usage

1. Interactive demo (Gradio)

The dataset demos need the sample sets under datasets/. Each of the first three demos auto-downloads its own samples from Hugging Face on first launch (the custom demo takes your own upload instead), or you can prefetch them all manually:

hf download GwanHyeong/InSpace --include "datasets/*" --local-dir .

Four demos share the same pipeline and UI, differing only in the input dataset:

# ERP-FRONT test scenes (fully-supervised: GT mesh + boxes)
python demo/app_inspace_erp_front.py --port 7860

# Structured3D scenes (windows/doors opening test; full / empty variants)
python demo/app_inspace_structured3d.py --port 7861

# ReplicaPano real-scan panoramas
python demo/app_inspace_replicapano.py --port 7862

# Your own 360° panorama (upload an ERP image + its depth map)
python demo/app_inspace_custom.py --port 7863

Each demo walks through every stage, input panorama to layout (PSG) to coarse geometry (CSG) to 3D boxes to textured mesh.

2. Batch inference

eval/pipeline/eval_pipeline.py chains all stages end-to-end over the test set (or your own data) and writes meshes and visualizations:

# Random-noise start, GT boxes (baseline)
python eval/pipeline/eval_pipeline.py \
    --data_dir datasets/ERP_3D_FRONT_test \
    --noise_mode random --bbox_mode gt --max_samples 8

# Layout-guided (SDEdit) start with predicted boxes (main setting)
python eval/pipeline/eval_pipeline.py \
    --data_dir datasets/ERP_3D_FRONT_test \
    --noise_mode sdedit --sdedit_alpha 0.5 --bbox_mode predicted

Key flags: --noise_mode {random,sdedit}, --sdedit_alpha, --bbox_mode {gt,predicted}, --layout_mode, --max_samples, --skip_existing. Texture is always generated. The seven convenience wrappers eval/pipeline/run_*.sh cover the full flow (see eval/README.md).

Per-sample outputs under evals/stage12_pipeline/{config}/{scene}/{room}/: latents (shape_latent.npz, texture_latent.npz, bboxes.npz), meshes (meshes/{scene,layout}.glb, meshes/assets/*.glb), and render grids in vis_pred/ (plus vis_concat/ with --save_concat).

🗂️ Data Preparation

InSpace is trained on ERP-FRONT, a paired ERP-Image-to-3D indoor scene dataset built on 3D-FRONT (26.5K training and 2.5K test ERP-image-mesh pairs), stored under datasets/ERP_3D_FRONT and datasets/ERP_3D_FRONT_test. Raw 3D-FRONT rooms are converted to the O-Voxel representation and conditioning inputs by the scripts in data_toolkit/erp/ (steps 1 to 10: mesh/PBR dumps, dual-grid O-Voxels, shape/PBR/SS latents, cubemap rendering, depth-lifted voxels), run as two parallel tracks: scene+assets (*_erp) and layout (*_layout_wo_ceiling). See data_toolkit/erp/README.md for the full step-by-step guide, and the expected on-disk layout in datasets/README.md. Optional or experimental variants are kept under data_toolkit/erp/extra/.

🏋️ Training

Training is driven by train.py (see --help for distributed flags). Bash wrappers set the config and output dir for each stage:

# Coarse scene geometry (sparse structure)
bash scripts/train/stage1_ss.sh

# 3D bounding-box estimator (CenterPoint)
bash scripts/train/bbox.sh

# Asset-aware shape generation
bash scripts/train/stage2_shape.sh              # stage2_shape_resume.sh to resume
# Asset-aware texture generation
bash scripts/train/stage2_texture.sh            # stage2_texture_resume_weighted.sh for large-room oversampling

The *_weighted script enables room-area weighted sampling (oversamples large rooms, see --sampler). Training writes to results/…, while released or served weights live in ckpts/….

📊 Evaluation

While Batch inference above is a single ad-hoc run, reproducing the paper's metrics is a three-step flow that wraps it with GT reconstruction and metric computation, all driven by eval/pipeline/ (see eval/README.md for the full guide):

bash eval/pipeline/run_recon_gt.sh                       # 1. GT reconstruction (once) → evals/gt_recon/
bash eval/pipeline/run_sdedit_predicted.sh               # 2. batch inference (shape)
bash eval/pipeline/run_texture_sdedit_predicted_0.5.sh   #    + texture (α sweep: 0.3 / 0.5 / 0.7)
bash eval/pipeline/run_metrics.sh                        # 3. metrics over all configs

• eval/pipeline/: end-to-end runner (eval_pipeline.py), GT reconstruction (recon_gt.py), and metrics (compute_metrics.py) — 3D voxel-IoU / Chamfer / F1 and 2D PSNR / SSIM / LPIPS, at scene and asset level (assets matched by voxel centroid).
• eval/stage1/: component evals — coarse-structure (SS) ablation (step1_generate.py → step2_metrics_vis.py → step3_metrics_vs_scene_properties.py) and the 3D bbox estimator (bbox_inference_centerpoint*.py).
• eval/viewers/: the interactive GT-vs-pipeline viewer (gt_vs_pipeline_unified_viewer.py) and the paper-figure grid generator (create_stage12_comparison.py); superseded viewers (incl. the turntable-video variant) are under eval/viewers/legacy/.

📁 Repository Layout

InSpace/
├── trellis2/             # core model: TRELLIS.2 + InSpace ERP / asset-aware extensions
├── data_toolkit/erp/     # dataset preprocessing (mesh to O-Voxel, latents, cubemap, depth voxels)
├── configs/{gen,bbox,scvae}/   # training configs  (legacy bbox under configs/bbox/legacy/)
├── demo/                 # interactive InSpace scene demos (app_inspace*.py)
├── eval/                 # evaluation: pipeline/ (runner + metrics), stage1/, viewers/
├── scripts/train/        # training launch scripts (+ legacy/ for old bbox variants)
├── ckpts/                # downloaded weights go here (gitignored)
├── datasets/             # (stubs) place ERP_3D_FRONT / _test here
├── o-voxel/              # vendored O-Voxel package
├── tools/                # standalone analysis / visualization utilities
├── figures/              # doc images 
└── train.py              # training entrypoint

🙏 Acknowledgements

InSpace is built on top of TRELLIS.2 (O-Voxel, SC-VAE, flow-matching DiTs), is inspired by OmniPart for part-aware generation, and uses DA2 for monocular ERP depth. We thank the authors of these projects. See o-voxel/, FlexGEMM, and CuMesh for the underlying high-performance packages.

💰 Funding & Support

This work was supported by the Korea Planning & Evaluation Institute of Industrial Technology (KEIT) and the Ministry of Trade, Industry & Resources (MOTIR) of the Republic of Korea (RS-2024-00417108), and by the Institute for Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2021-II211381, Development of Causal AI through Video Understanding and Reinforcement Learning, and Its Applications to Real Environments).

⚖️ License

This model and code are released under the MIT License. InSpace is built upon TRELLIS.2 (MIT License, Copyright (c) Microsoft Corporation).

This project is implemented for non-commercial, academic research purposes only.

Please note that certain dependencies operate under separate license terms:

• nvdiffrast: Utilized for rendering generated 3D assets. This package is governed by its own License.
• nvdiffrec: Implements the split-sum renderer for PBR materials. This package is governed by its own License.

Both packages are licensed for non-commercial research and evaluation purposes only, with commercial rights reserved to NVIDIA Corporation and its affiliates. InSpace does not incorporate source code from these packages, but they are required at runtime; accordingly, use of InSpace is limited to research and evaluation purposes.

📚 Citation

InSpace has been accepted to ECCV 2026. If you find our work useful, please cite:

@article{koo2026inspace,
  title   = {InSpace: Structure-Aware 3D Indoor Scene Generation from a Single 360{\deg} Image},
  author  = {Koo, Gwanhyeong and Kim, Hyunsu and Kim, Youngji and Lee, Taejae and
             Lim, Siwoo and Yoon, Sunjae and Yeon, Suyong and Yoo, Chang D.},
  journal = {arXiv preprint arXiv:2607.03990},
  year    = {2026}
}