README.md

---
language:
- en
tags:
- 3d
size_categories:
- 100M<n<1B
---

# Imaginarium: Vision-guided High-Quality 3D Scene Layout Generation
**SIGGRAPH ASIA 2025 & ACM Transactions on Graphics (TOG)**

## 📖 Introduction

**Imaginarium** is a novel vision-guided 3D layout generation system that addresses the challenges of generating logically coherent and visually appealing customized scene layouts. We employ an image generation model to expand prompt representations into images, fine-tuning it to align with our high-quality asset library. We then develop a robust image parsing module to recover the 3D layout of scenes based on visual semantics and geometric information, optimizing the scene layout using scene graphs to ensure logical coherence.

## 📦 Data Preparation

The 3D scenes and asset dataset are hosted at [🤗 HiHiAllen/Imaginarium-Dataset](https://huggingface.co/datasets/HiHiAllen/Imaginarium-Dataset), and the derived dataset is hosted at [🤗 binicey/Imaginarium-3D-Derived-Dataset](https://huggingface.co/datasets/binicey/Imaginarium-3D-Derived-Dataset).

### 1. 3D Scenes and Asset Dataset Downloads

Choose the appropriate package based on your needs:

#### Plan A: Full 3D Scene Layout Dataset (Research)
For full access to Blend source files, RGB renders, instance segmentation, bounding boxes, depth maps, and meta-info (captions, scene graphs, object poses), download:
- `imaginarium_3d_scene_layout_dataset_part1.tar.gz`
- `imaginarium_3d_scene_layout_dataset_part2.tar.gz`
- `imaginarium_3d_scene_layout_dataset_part3.tar.gz`
- `imaginarium_3d_scene_layout_dataset_part4.tar.gz`

**Structure (e.g., bedroom_01):**
```text
bedroom_01/
  ├── bedroom_01.png
  ├── bedroom_01.blend
  ├── bedroom_01_bbox_overlay.png
  ├── bedroom_01_depth_vis.png
  ├── bedroom_01_depth.npy
  ├── bedroom_01_detect_items.pkl
  ├── bedroom_01_meta.json
  └── bedroom_01_segmentation.png
```

#### Plan B: Flux Fine-tuning Data Only
If you only need data for fine-tuning Flux (RGB images & meta-info), download:
-  `flux_train_data.tar.gz`

#### Plan C: Running Imaginarium (Inference)
To run the algorithm using our provided weights, you need the 3D Asset Library and metadata:
- `imaginarium_assets.tar.gz` (3D Models)
- `imaginarium_assets_internal_placement_space.tar.gz` (Internal Placement Spaces Info)
- `imaginarium_asset_info.csv` (Metadata)
- `background_texture_dataset.tar.gz`（Background Texture Dataset）
- *(Optional)* `imaginarium_asset_info_with_render_images.xlsx` (Visual Reference)