---
title: 'SpatialEdit-Bench: Spatial Editing Evaluation'
url: https://www.emergentmind.com/topics/spatialedit-bench
type: topic
---

# SpatialEdit-Bench: Spatial Editing Evaluation

SpatialEdit-Bench is a geometry-aware benchmark explicitly designed for evaluating the fine-grained spatial editing capabilities of image generative models. It assesses not only the perceptual plausibility of editing outputs but, critically, their adherence to prescribed geometric transformations, operationalized via object-centric and camera-centric manipulations. The benchmark is structured to diagnose an editor's grasp of "where" and "how" edits by verifying both photorealism and implementation of intended 3D changes, addressing a major shortcoming of previous benchmarks which lack sensitivity to spatial precision [2604.04911].

## 1. Task Definition and Axis Structure

SpatialEdit-Bench formalizes spatial image editing along two orthogonal axes: object-centric manipulation and camera-centric view control. Each axis encompasses specific, parameterizable tasks:

- **Object-centric manipulation:**
  - *Translation*: Relocating a designated object into a target bounding-box region within the image.
  - *Scaling*: Resizing the object to fit a given rectangle, effecting changes in both scale and placement.
  - *Rotation*: Reorienting the object to one of eight canonical views (e.g., front, front-right, right, etc.).

- **Camera-centric view control:**
  - *Yaw*: Applying horizontal rotation (orbiting) of the virtual camera in discretized 45° steps.
  - *Pitch*: Adjusting vertical tilt in discretized 15° increments.
  - *Zoom*: Modifying camera distance to perform zoom-in/zoom-out operations toward a focus object, altering both scale and perspective.

This structure enables isolated and compositional evaluation of an editor's capabilities in reproducing precise spatial and viewpoint changes.

## 2. Evaluation Metrics

SpatialEdit-Bench introduces a dual-metric suite to quantify both semantic plausibility and geometric accuracy:

- **Perceptual Plausibility (LPIPS):**
  - Adopted from learned perceptual similarity metrics, the LPIPS distance is computed as
    $$
    d_{\mathrm{LPIPS}}(x,y) = \left\lVert \phi(x) - \phi(y) \right\rVert_2,
    $$
    where $\phi(\cdot)$ denotes a pre-trained deep feature extractor, typically VGG. Lower values reflect increased perceptual similarity, serving as a baseline for plausibility.

- **Geometric Fidelity:**
  - *Viewpoint Error (VE)*: Integrates translation and rotation errors in predicted camera pose, measuring deviation from ground-truth in $\mathrm{SE}(3)$. This is defined as:
    $$
    \epsilon_{xyz} = \frac{\left\|\mathbf{C}_{\mathrm{pred}} - \mathbf{C}_{\mathrm{gt}}\right\|_2}{\left\|\mathbf{C}_{\mathrm{gt}} - \mathbf{C}_{\mathrm{src}}\right\|_2 + \varepsilon},
    $$
    $$
    \epsilon_{\mathrm{rot}} = \frac{1}{90} d_{\mathrm{geo}}(R_{\mathrm{pred}}, R_{\mathrm{gt}}),
    $$
    $$
    \mathrm{VE} = \frac{1}{2}(\epsilon_{xyz} + \epsilon_{\mathrm{rot}}),
    $$
    with the rotational distance $d_{\mathrm{geo}}$ given by the matrix trace formula.

  - *Framing Error (FE)*: Uses object detectors to verify that salient elements maintain correct image-plane layout post-edit. This metric incorporates angular misalignment and zoom-direction correctness. FE is computed as:
    $$
    \mathrm{FE} = \frac{1}{2}(\epsilon_{\mathrm{rag}} + \epsilon_{\mathrm{zde}}),
    $$
    where $\epsilon_{\mathrm{rag}}$ measures angular deviation between detected boxes and $\epsilon_{\mathrm{zde}}$ encodes correct zooming directionality.

- **Object Sub-task Scores:**
  - *Moving Score (MS)* and *Rotation Score (RS)* leverage both IoU of bounding boxes and VLM-derived object/view consistency. The overall object score takes their mean.

A summary of the principal metrics is given below:

| Metric                       | Purpose                       | Mathematical Definition/Notes                   |
|------------------------------|-------------------------------|------------------------------------------------|
| LPIPS                        | Perceptual similarity         | $d_{\mathrm{LPIPS}}(x,y)$                      |
| Viewpoint Error (VE)         | Camera pose accuracy          | Combines translation/rotation errors           |
| Framing Error (FE)           | Image-plane layout correctness| Considers bounding box alignment, zoom         |
| MS, RS, Object Overall Score | Object manipulation accuracy  | VLM and detector-based, meaned for summary     |

These metrics are designed to reveal not only visual believability but the edit's fidelity to the instructed geometric intent.

## 3. Ground-Truth Transformations and Representation

All edits in SpatialEdit-Bench are grounded in explicit 3D transformations:

- **Object-centric edits** are formalized as rigid-body transforms in $SE(3)$:
  $$
  T_{\mathrm{obj}} = \begin{pmatrix} R_{\mathrm{obj}} & t_{\mathrm{obj}} \\ \mathbf{0}^\top & 1 \end{pmatrix} \in SE(3),
  $$
  where $R_{\mathrm{obj}} \in SO(3)$ (rotation) and $t_{\mathrm{obj}} \in \mathbb{R}^3$ (translation). Scaling is ensured via uniform scale in the rendering process.

- **Camera-centric edits** are parameterized by camera extrinsics $(R, t)$. Ground-truth, source, and predicted cameras are encoded, enabling rigorous geometric comparison using both direct pose error:
  $$
  E_{\mathrm{pose}} = \|R_{\mathrm{pred}}^T R_{\mathrm{gt}} - I\|_F + \|t_{\mathrm{pred}} - t_{\mathrm{gt}}\|_2,
  $$
  and the normalized errors described in VE.

- **2D Image-Plane Projections**: For evaluating layout-specific metrics like FE, precise 2D bounding box projections are calculated from known 3D geometry.

This level of control and annotation enables unambiguous evaluation of spatial edits in a physically grounded manner.

## 4. The SpatialEdit-500k Synthetic Dataset

Benchmarking with precise spatial supervision necessitates large, annotated datasets. SpatialEdit-500k addresses this via Blender-powered synthetic generation:

- **Object-centric pipeline**:
  - $\sim$10K GLB assets from TexVerse form the object pool.
  - Each asset is rendered in a canonical front view; VLM filtering ensures correct pose.
  - Eight canonical viewpoints (via discrete in-place rotation), random 2D translations and 3D uniform scaling are applied.
  - Semantic segmentation is performed using SAM3.
  - Foregrounds are composited with backgrounds generated by a text-to-image model conditioned on object class.
  - 3D bounding box projections produce exact box labels.

- **Camera-centric pipeline**:
  - $\sim$200 curated 3D scenes, with $N_{\mathrm{scene}}^{\mathrm{target}}$ salient focus objects per scene.
  - Cameras are sampled across yaw, pitch, distance grid; object/truncation errors are filtered via YOLO and VLM.
  - Camera parameters $(\theta_y, \theta_p, d)$ are recorded, and source-target pairs are formed with known transformation deltas.
  - Templated and natural-language instructions are auto-generated for each edit.

The dataset ultimately comprises approximately 500,000 image pairs, systematically balanced across all six benchmark sub-tasks.

## 5. Benchmark Evaluation Protocol

SpatialEdit-Bench defines a comprehensive evaluation protocol to ensure objectivity and rigor:

- **Hold-out splits**: Specific objects and scenes are reserved for test to eliminate memorization or overfitting.
- **Per sub-task evaluation**:
  - Editors receive $(I_{\mathrm{src}},\,\text{instruction})$ and must render $I_{\mathrm{pred}}$.
  - Object-centric tasks are scored with MS and RS, averaged for the object overall score (higher is better).
  - Camera-centric tasks are scored with VE and FE, averaged for camera overall error (lower is better).
- **Model ranking**: By default, models are ranked by descending object overall and ascending camera error; a composite normalized score is also supported.

This protocol facilitates fair, reproducible comparison of spatial editing across systems.

## 6. Empirical Benchmarks and Baseline Performance

SpatialEdit-16B serves as a reference model, trained on the SpatialEdit-500k dataset. Comparative performance with existing methods is as follows:

| Task                 | SpatialEdit-16B | LongCat | Difference        |
|----------------------|-----------------|---------|-------------------|
| Object moving score (MS)   | 0.673           | 0.373   | +0.300            |
| Object rotation score (RS) | 0.632           | 0.505   | +0.127            |
| Viewpoint error (VE)       | 0.243           | 0.802   | –0.559            |
| Framing error (FE)         | 0.527           | 0.684   | –0.157            |
| Object overall             | 0.653           | —       | —                 |
| Camera overall error       | 0.385           | —       | —                 |

SpatialEdit-16B achieves strong semantic plausibility while substantially improving on geometric fidelity, especially on object manipulation and precise viewpoint control [2604.04911]. This suggests that SpatialEdit-Bench is capable of differentiating between general attention-based inpainting and true spatial understanding.

## 7. Significance and Diagnostic Scope

SpatialEdit-Bench advances the evaluation of image editing models by enabling joint assessment of semantic believability and spatial manipulation fidelity. Its geometry-aware supervision, task granularity, and rigorous metric design allow researchers to distinguish superficial plausibility from genuine spatial competence. A plausible implication is that, as editors trained and evaluated via SpatialEdit-500k approach lower Viewpoint/Framing Errors and higher object scores, real-world editing pipelines can attain confidence not only in visual plausibility but also in faithful execution of precise geometric instructions [2604.04911].

Source: https://www.emergentmind.com/topics/spatialedit-bench