Thinking in Boxes

A structured interface
for 3D image editing

Text and 2D-conditioning interfaces provide weak, ambiguous control over spatial transformations in image editing — particularly under large object motions and camera changes. Prior work has used 3D primitives such as boxes, but only as loose conditioning signals indicating approximate object location rather than specifying the transformation.

We instead use 3D boxes as structured specifications: the user provides the input and output boxes of the edit, casting editing as a well-posed geometry problem. This "thinking in boxes" interface, where each box face is color-coded to convey 3D orientation, gives precise control over translation, rotation, scaling, and viewpoint changes in real images while preserving scene and object identity, and recovering previously unseen object regions. To ground transformations in scene appearance, we introduce a depth-aligned planar floor as a global reference frame, shaded with depth-aware cues.

Conditioned on this structure, an image generator produces consistent results under large transformations. Trained in two stages — on synthetic multi-object scenes and a small set of real-world videos from Objectron — the system generalizes to complex, in-the-wild real images. Our method operates directly on real photographs and substantially outperforms recent state-of-the-art methods on large 3D edits.

How it works

Editing pipeline. (1) Users provide a real source image and (2) fit 3D boxes to the objects within the scene using a point-and-click interface. (3) The boxes can be manipulated in 3D space, allowing for scaling, rotation, translation, and camera moves. Both the source and target box layouts are projected into 2D and serve, alongside the source image, as inputs to an image editing model. (4) The model generates an edit that respects the underlying scene geometry and follows the user's layout.

Try the box editor

A bite-sized preview of our editor. Pick a source scene, choose an edit operation, play around with the controls and the right panel shows the corresponding generated result. For full control over the boxes, use the Demo provided in the project repository.

📦 Thinking in Boxes — Interactive Demo Pick a scene

TTranslate RRotate SScale CCamera

Box Scene Fit Boxes

Source

Drag to translate Drag the slider — or drag directly inside the Box Scene.

Scene A · Owl cup B · Green car C · Coming soon D · Coming soon Pick a mode (T/R/S/C) then drag inside the box scene or the slider

Box face color code

Front

Back

Top

Bottom

Left

Right

Dataset & training pipelines

Our model is trained in two stages:
I) A large synthetic dataset rendered in Blender and
II) Fine-tuning on real-world Objectron videos.

Data Synthetic Dataset Pipeline

Synthetic data generation. Multi-object scenes are assembled by sampling assets from ObjaverseXL, HDRIs, and material libraries, then placed in Blender. We render paired RGB images and 3D-primitive box layouts for source/target edit configurations.

Train Two-Stage Training Pipeline

Training. We condition FLUX-Kontext on (source image, source boxes, target boxes). A learned transformation 𝒯 maps source latents to target latents under guidance of the projected box layouts.

Training at a glance

A snapshot of the data and recipe used to produce the model behind every result on this page.

Qualitative results
across edit types

Pick a set A–C or Baselines to browse the full grid of results.
Click to play its image → source box → target box → generation transition.

🖼️ Result Gallery Hover to preview · Click ▶ to play

A B C Baselines

BibTeX

If you find our work useful, please consider citing:

@misc{bhat2026thinkingboxes3dediting,
  title         = {Thinking in Boxes: 3D Editing in Real Images Made Easy},
  author        = {Pradhaan S Bhat and Naveen Chandra R and Rishubh Parihar and Vaibhav Vavilala and R. Venkatesh Babu and D. A. Forsyth and Anand Bhattad},
  year          = {2026},
  eprint        = {2606.20556},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2606.20556}
}