SpatialEdit lets you move objects and rotate the camera in 2D photos

Anyone who has tried asking an AI to edit a photo knows the pain. Tell a modern model to "scoot the dog two inches to the left" or "tilt the camera down slightly," and it usually re-rolls the dice entirely. The dog's breed changes, the background distorts, and the original shot is ruined.

A new image editing model called SpatialEdit dropped as a fully open-source project to tackle this exact geometric bottleneck. Instead of guessing through fuzzy text prompts, it lets you manipulate 2D photos as if they were live 3D scenes.

Teaching 2D diffusion models the rules of 3D

The fundamental problem with spatial image editing has always been the lack of precise ground-truth data. To crack this, the researchers built SpatialEdit-500k, a synthetic training dataset generated with a controllable Blender pipeline.

By systematically rendering objects across varying environments with exact camera trajectories, they created 500,000 paired examples with true geometric coordinates for both object shifts and camera pans.

Built on top of that synthetic foundation is SpatialEdit-16B. The model's interface allows users to tweak the exact placement and rotational angle of individual subjects, as well as adjust the virtual camera's yaw, pitch, and zoom.

Alongside the model, the team introduced SpatialEdit-Bench, a dedicated benchmark measuring both perceptual plausibility and geometric fidelity. In side-by-side evaluations, SpatialEdit outperformed existing tools like nano banana and cradedit. While community solutions like dedicated LoRAs for Qwen Image Edit have attempted spatial control before, testers noted that SpatialEdit is currently the only model consistently pulling off fine-grained geometric shifts.

Everything is open, but you will need serious hardware

The most refreshing part of the release is how open the team has been. They did not just share a flashy demo video; they released the complete codebase, the full 500k training dataset, and the model weights.

The installation instructions walk users through local deployment from scratch.

The catch is local accessibility. The model weights weigh in at roughly 32 GB, meaning you will need a relatively beefy, high-end GPU with serious VRAM to run it comfortably on your own rig.

Why it matters

We think SpatialEdit's smartest move was stopping the hopeless attempt to explain Euclidean space through plain English prompts. Injecting 500k Blender-generated ground truths gives the diffusion backbone an actual spatial compass rather than a guessing game.

For commercial creators, visual designers, and product photographers, this bridges a massive workflow gap. Moving a product slightly to the left or adjusting an angle previously required re-rendering the entire 3D asset or spending hours masking in Photoshop. Bringing parametric control directly to 2D generations makes AI image generation feel less like an unpredictable slot machine and more like an actual professional tool.

It also puts pressure on mainstream image generators to upgrade their spatial manipulation suites. The main hurdle right now is the 32 GB footprint, which locks out hobbyists without professional GPUs, plus the lingering question of how well Blender-trained synthetic transforms hold up in chaotic real-world photography. Even so, publishing the weights alongside the entire training dataset gives the open-source community a massive foundation to build on.