Scene Dreamer: Generate unbounded 3D scenes from 2D images
What is Scene Dreamer?
SceneDreamer is an unconditional generative model for unbounded 3D scenes. Starting from simplex noise and a style code, it synthesizes large-scale 3D landscapes that can be viewed along a free camera trajectory with 3D consistency and well-defined depth. The framework is trained end-to-end on in-the-wild 2D images only, with no 3D annotations required. The method combines three parts: a bird's-eye-view scene representation made of a height field and a semantic field, a generative neural hash grid that parameterizes latent features from 3D positions and scene semantics, and a style-modulated neural volumetric renderer learned through adversarial training. The project page links a paper, video, code and demo, and the work was published in IEEE TPAMI in 2023.
Focus areas: Image to 3D.
Who is Scene Dreamer for?
- 3D graphics and vision researchers
- Machine learning engineers exploring generative 3D models
- Technical artists experimenting with scene synthesis
Is Scene Dreamer worth it?
A well-documented research model for unbounded 3D scene generation, best suited to researchers who will run the code themselves
Where Scene Dreamer is strong
- Generates unbounded 3D scenes with free camera movement
- Trains from 2D image collections without 3D annotations
- Published in IEEE TPAMI 2023 with paper, code and demo links
Where it falls short
- No pricing or hosted product information is published
- Output quality and setup details depend on the linked code and paper, not the page
How do you get started with Scene Dreamer?
- 01Open the project page at scene-dreamer.github.io
- 02Read the paper and watch the video for the method overview
- 03Follow the Code and Demo links to run the model
Scene Dreamer: frequently asked questions
What does SceneDreamer actually generate?
It generates large-scale, unbounded 3D landscapes from random noise, with 3D consistency, well-defined depth and a free camera trajectory.
Does it need 3D training data?
No. The framework is learned from in-the-wild 2D image collections only, without any 3D annotations.
How much does it cost?
No pricing is published on the project page, and the pricing URL returns a 404, so cost is unknown.
Who made it and where was it published?
Zhaoxi Chen, Guangcong Wang and Ziwei Liu of Nanyang Technological University published it in IEEE TPAMI in 2023.
Scene Dreamer alternatives
The closest tools in our directory, matched on category, focus areas and audience.
| Tool | Best for | Free plan | Starting price | Focus areas |
|---|---|---|---|---|
| Scene Dreamer | Researchers studying generative 3D scene synthesis from 2D images | Unknown | — | Image to 3D |
| GET3D (Nvidia) | Researchers and developers exploring image-to-3D shape generation | Unknown | — | Image to 3D |
| Farm3D | Researchers reconstructing articulated 3D animals from single images | Unknown | — | — |
| Genie 3 World | Create interactive 3D AI worlds from text prompts with real-time physics. | Unknown | — | — |
| TRELLIS 3D AI | Turning reference images into downloadable 3D assets | Yes | $16/mo | Image to 3D |
| ImgTo3D | Early-stage 3D concepts from existing photos or illustrations | No | $19.90/mo | Image to 3D |
| PicToMesh | Turning photos and illustrations into downloadable 3D assets | No | — | Image to 3D |
| Stable Fast 3D(SF3D) | Fast single-image 3D asset creation for games, VR and e-commerce | Unknown | — | Image to 3D |
| Fast3D | Quick 3D asset creation from prompts or reference images | Yes | $7/mo | Image to 3D, Text to 3D |
| Sloyd | Game devs, designers and 3D printing users needing fast 3D assets | Yes | $15/mo | Image to 3D, Text to 3D |
| 3D AI Studio | Game developers and creators needing fast 3D assets | Yes | — | Image to 3D, Text to 3D |
