SpatialEdit Repositions Reality In Static Images

By vramkickedin | | 2 min read
A large cube constructed from white thin glowing white wireframe lines with a camera that is suspended next to it and bokeh particles.

SpatialEdit is an open source research model that performs precise, geometry-driven changes to static images. It moves objects, rotates items, and shifts camera angles while keeping the original scene and subjects intact.

Created by Yicheng Xiao and his research team, the project addresses the limitations of existing editing tools that often struggle with accurate spatial control. Local operators can run the model directly to generate custom visual layouts without relying on cloud servers.

Model Size: 32.8GB & VRAM GPU: requirements vary

Handling precise image geometry and camera shifts

  • Follows direct written commands to reposition objects or adjust viewing angles.
  • Preserves original lighting and subject identity during structural changes.
  • Trained on a synthetic dataset built with Blender for accurate spatial ground truth.
  • Includes a complete testing suite that checks visual realism and geometric accuracy.
  • Supports generating smooth camera transition videos from a single starting frame.

Visual designers working on tight deadlines may find the local installation workflow useful for quickly testing layout variations without uploading sensitive assets. Small studios can use the included benchmark tools to verify that specific movements match the original creative brief before finalizing outputs.

Setup requirements and current limitations

Running the software locally requires updating several configuration files to match your specific directory paths before generating results. The setup depends on downloading external checkpoints for video generation, and it needs a graphics environment that supports recent CUDA and PyTorch versions.

Users should note that some internal file paths remain as placeholders in the downloaded scripts.

"Current models are insufficient for fine-grained spatial manipulations, motivating a dedicated assessment suite,"

noted the team in a technical paper. Adjusting the configuration variables ensures the system correctly locates the supporting modules during execution.

Explore the SpatialEdit project files, review the technical paper, or access the pre-trained 16B model checkpoint.