UniGeo lets you edit images by moving the camera at exact angles

If you have ever tried asking an AI image editor to shift the camera just a little to the right, you know the frustration.

Most existing models do not actually move a virtual camera. They just regenerate the picture from scratch, hallucinating a different face or adding an extra leg to the table in the process.

Researchers introduced UniGeo, a camera-controllable image editing framework built to fix this exact issue. It turns a single 2D image into a 3D geometric representation, letting you specify exact camera angles and trajectories just like adjusting a virtual tripod.

Camera control with degree-level precision

UniGeo is built around precise, quantitative motion rather than vague descriptive prompts.

Instead of saying look from the side, you can feed it concrete numbers. The camera does not need to stay locked on a central subject; you can freely pan, tilt, or shift anywhere in 3D space.

Even better, you can chain multiple camera movements together in a single prompt. For instance, you can tell the model to pan 16 degrees to the left and then tilt up 7 degrees. UniGeo follows that continuous trajectory and renders the final target view accordingly.

Compared to general-purpose editing tools like Nano Banana or GPT Image 2, having real numerical control over viewing angles makes novel view editing feel predictable for once.

Borrowing priors from video models

Previous attempts at camera-guided image editing struggled with what researchers call geometric drift and structural degradation.

Earlier methods mostly relied on image diffusion models operating across discrete viewpoint jumps. Because geometric guidance was only injected at fragmented points, the moment you moved the camera further, the 3D scene fell apart.

UniGeo takes a smarter architectural approach by borrowing the continuous viewpoint priors of video models, backing them up with unified geometric guidance across three critical layers.

At the representation level, it reconstructs 3D point clouds and injects decoupled geometric references. At the architecture level, it uses geometric anchor attention to align multi-view features. At the loss level, it adds trajectory-endpoint geometric supervision to strictly enforce structural fidelity at the final frame.

Benchmarks show it handles both narrow and wide camera sweeps without distorting objects. The project page currently lists the model as coming soon, signaling that an open release is likely around the corner.

Why it matters

For product designers, digital marketers, and concept artists, this could turn single-view photos into flexible assets. If a client wants to see a product slightly from above or from an oblique angle, you no longer have to build a full 3D CAD model or roll the dice on endless random generations.

Industry-wide, UniGeo marks a clear shift toward deterministic generation. Raw visual fidelity is largely a solved baseline; the new battleground is precision control. By using video models to handle smooth transitions and point clouds to anchor the 3D geometry, the framework offers a very practical blueprint for true view synthesis.

That said, running 3D reconstruction together with video diffusion is computationally heavy. We will also need to see how well it fills in heavily occluded areas when the camera takes dramatic leaps once the model weights are actually in the wild.