Bottom line: This is an account-gated scene-reconstruction toy, not “one image becomes an infinite map.” You drop in a room photo; it tries to return a small space you can orbit, with objects you can nudge. The doorway is often visible first; finished results need a login and may spend credits.
One-minute overview
- What you do: Split a scene photo into movable foreground props and a background you can look around.
- First move: Open the Hyper3D Rodin workspace, switch to WorldGen, and inspect the entry. Log in only when you are ready to generate.
- Where the AI sits: The company news release says editable objects are separate Rodin meshes, while the background uses Gaussian splats; the research behind it is CAST.
- Is there an ending? No campaign. One run produces an editable small scene, not an open world.
- Language: The workspace is English. You do not need a modelling app to try the first photo.
Why it’s interesting
Most image-to-3D tools lift out one hero object. WorldGen is trying something else: turn a still frame into a tiny stage you can restage. The 9 September 2026 news release says you upload a scene image, then either let the system identify main objects or draw boxes around what should be generated. Those assets can be edited, replaced, and moved on their own.
The hybrid representation sounds more technical than it needs to:
- Foreground: things you want to move, such as a sofa or lamp, as separate meshes.
- Background: walls, windows, and floor as a Gaussian-splat environment you can look around. That preserves detail. It does not mean infinite streets.
Hyper3D traces the research to CAST (component-aligned reconstruction from one RGB image), which received a SIGGRAPH 2025 Best Paper. WorldGen is the product-shaped door. The news release also describes estimated contact, support, pose, scale, and collision. That is the company’s claim about the system. This site did not re-measure those quantities on your photos, and the result is not a survey tool or a home-buying reference.
It is still limited to what one picture can imply. Earlier verification recorded that this is not a one-image infinite open world. The news release lists games, film, robot simulation, and spatial computing as intended uses. Those pipelines were not re-tested here. The product pitch we can stand behind is “scene-level,” not “another figurine.”
How to try it
First: look at the door, then decide whether to generate
- Open the original link on this page, enter the Hyper3D Rodin workspace, and switch to WorldGen. The entry is often visible before login.
- Generating needs an account. The public site says you can start with free credits and without a subscription. Quotas, Credits, and plans are whatever the live page shows that day.
- Pick a clear indoor photo with little occlusion. A living room where the sofa, table, and walls are obvious beats a night street.
- Upload when prompted and choose which objects should be generated on their own. Official copy allows automatic detection or drawing boxes.
- Skip crowds, dense streets, and aerial city views on a first try. Generation takes time and is not guaranteed to drop in as a finished game level.
This site checked the public entry and official statements. It did not upload your images or spend your credits. A desktop browser is better for inspecting 3D results.
Workspace words, as a map rather than a promise
Labels can change. These come from earlier verification and the news release, so you can recognise the step you are on.
| You may see | When it matters |
|---|---|
WorldGen | Switch to this inside the Rodin workspace; do not stay on single-asset generation. |
Upload a scene image | Use an indoor photo you have the right to use. |
Pick objects | Tick only the props you actually want to move. |
| Auto-detect / draw boxes | Two official ways to choose objects. |
| Separate meshes | Foreground props you can nudge or export. |
| Gaussian splat / 3DGS environment | The look-around background, not an infinite map. |
First small goal: one living-room photo, understood as something you can take apart
| Step | Do this | Success looks like |
|---|---|---|
| 1 | Find the WorldGen entry | You can see where a scene image would go |
| 2 | Log in and upload a low-occlusion indoor photo | The page accepts the image and asks you to pick objects |
| 3 | Select only one to three clear pieces of furniture | You did not tick the back wall or people outside |
| 4 | Wait for one generation | You can orbit, and at least one foreground object can be selected alone |
Stop there. Ask three questions: can I look around this room? can I move that lamp? did it inflate one photo into a city? The first two are the product. If the third is yes, the expectation is wrong, not the button.
Three ways to keep going
- Set dresser. Move the table, swap the lamp, drop the camera. The news release frames this as preview for film and spatial work; playing it as a fake room is enough.
- Disassembly test. Generate only the sofa, then run the same photo with the lamp and table included. See what extra foreground does to the background.
- Asset door. Official copy says results can continue in Blender, Unity, or Unreal. That is a capability claim, not a pipeline this site ran. If you do not model for a living, being able to orbit and nudge counts as having tried it.
Common stuck points
- Do not confuse this Hyper3D / Deemos WorldGen with other open-source projects that reuse the same English name. The entry here is the Rodin workspace.
- Do not upload other people’s photos, private pictures of real people, or interiors that show an address or ID.
- Aerial city shots ask the system for more than one image can imply.
- The door can be visible without a login; finished generation usually cannot.
- Robot-sim and film workflows in the news release were not re-tested here.
- Treat geometry as an estimate, not a measurement.
Requirements and sources
| Gate | Current picture |
|---|---|
| Cost | Freemium. Official site: start with free credits, no subscription required. Generation usually needs a login and may spend Credits. |
| Account | Optional to inspect; required to generate |
| Install | No special client; desktop browser preferred |
| Language | English workspace |
| Network | Required |
| Ending | None. One run is one editable small scene |
| Rights | You must have rights to the photo. Terms also say inputs and outputs may be used to operate, improve, and review the service |
| What this site checked | Public entry and official copy. No generation on your behalf |
Makers: Deemos / Hyper3D. Public demo record: 31 August 2026. Company news release: 9 September 2026. CAST’s SIGGRAPH 2025 Best Paper is the company’s academic citation, not an award issued by this site. The cover remotely cites a frame from the official demo and is not re-hosted here.
- Entry: [Hyper3D Rodin workspace](https://hyper3d.ai/workspace/rodin)
- Site: [hyper3d.ai](https://hyper3d.ai/)
- News release: [9 September 2026](https://www.prnewswire.com/news-releases/hyper3d-launches-worldgen-to-turn-single-images-into-editable-3d-scenes-302873750.html)
- Earlier item verification: 31 August 2026 demo; not an infinite open world; generation needs a login