SpatialEdit-500k: Synthetic Dataset for Spatial Edits

• SpatialEdit-500k is a synthetic dataset featuring 500k paired images with precise 3D spatial annotations and natural language edit instructions.
• It employs dual pipelines—object-centric with GLB-based synthetic rendering and camera-centric with systematic pose sampling—to ensure comprehensive spatial transformation benchmarks.
• The dataset supports model training and evaluation through detailed preprocessing protocols, performance metrics (e.g., Moving Score, Rotation Score), and fine-tuning strategies.

SpatialEdit-500k is a large-scale, synthetic dataset designed for the training and evaluation of fine-grained image spatial editing systems. It enables systematic benchmarking of geometry-driven image transformations, facilitating precise manipulation of both object layout and camera viewpoints via paired data, dense ground-truth 3D annotations, and structured instructions. The dataset addresses bottlenecks in scalable data generation and annotation for spatial editing, and is foundational for new models and evaluation protocols in this domain (Xiao et al., 6 Apr 2026).

1. Dataset Architecture and Generation Pipelines

SpatialEdit-500k comprises 500,000 paired images, generated end-to-end in Blender v2.9x, evenly split between object-centric and camera-centric edits. Two parallel data generation pipelines are employed, sharing a common rendering engine and fixed camera intrinsics.

• Object-Centric Pipeline:

Source assets are approximately 5,000 GLB models from the TexVerse library, spanning 50 categories (e.g., chairs, mugs, vases, tools). Each object is initially rendered with a three-point studio light at focal length $f = 50$ mm. The “front” view is verified for semantic correctness via Gemini-2.5; assets failing this check are discarded. For each object, eight canonical 3D rotations about the centroid are realized at yaw increments of 45°, holding pitch and roll at 0°. A secondary augmentation applies random translation $t \sim \mathrm{Uniform}(-0.1, 0.1)$ m and scale $s \sim \mathrm{Uniform}(0.8, 1.2)$, producing eight further variants per view. Foreground segmentation is conducted with SAM3. Backgrounds compatible with the front view are generated using Nano-Pro text-to-image models and composited via Poisson blending. The final object-centric set contains roughly 250,000 source-edited pairs, each involving translation, rotation, or scaling across canonical perspectives.

• Camera-Centric Pipeline:

The scene bank comprises approximately 1,000 photorealistic Blender scenes (600 indoor, 400 outdoor) with full static lighting and geometry. Up to five salient objects per scene are manually labeled as “focal.” For each, systematic pose sampling is performed over discrete yaw $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$, pitch $$\theta_p \in \{-30^\circ, -15^\circ, 0^\circ, +15^\circ, +30^\circ\}$$, and distance $d \in [2.0\,\mathrm{m}, 6.0\,\mathrm{m}]$ at 0.5 m intervals, while holding intrinsics fixed at $f = 35$ mm, with the principal point at image center. Each rendered view passes YOLO-v10 and QwenVL-30B filters to ensure object visibility (≥ 50%) and eliminate mesh intersections or unnatural framings. Valid source-target view pairs are drawn per object, recording relative pose deltas $(\Delta\theta_y, \Delta\theta_p, \Delta d)$ and associated templated natural language instructions. This results in approximately 250,000 camera-centric pairs that encode granular viewpoint modifications.

2. 3D Transformations in Object and Camera Space

All 3D transformations in SpatialEdit-500k are formulated in a right-handed coordinate system with Z-forward, X-right, and Y-up conventions. Edits are expressed as elements of $\mathrm{SE}(3)$, employing explicit rotation and translation formulations.

• Object-Centric Edits:
  • Rotation: Rigid rotations about the object centroid are parameterized by Euler angles $(\psi, \phi, \theta)$ as $t \sim \mathrm{Uniform}(-0.1, 0.1)$0, where each $t \sim \mathrm{Uniform}(-0.1, 0.1)$1 denotes standard rotation matrices about axis $t \sim \mathrm{Uniform}(-0.1, 0.1)$2.
  • Translation: Offsets $t \sim \mathrm{Uniform}(-0.1, 0.1)$3 sampled independently from $t \sim \mathrm{Uniform}(-0.1, 0.1)$4 m.
  • Scaling: Isotropic scale $t \sim \mathrm{Uniform}(-0.1, 0.1)$5 with $t \sim \mathrm{Uniform}(-0.1, 0.1)$6.
• Camera-Centric Edits:
  • Camera extrinsics: For a focus location $t \sim \mathrm{Uniform}(-0.1, 0.1)$7, camera center is given by $t \sim \mathrm{Uniform}(-0.1, 0.1)$8, with world-to-camera rotation $t \sim \mathrm{Uniform}(-0.1, 0.1)$9.
  • Relative transformations: For paired poses, $s \sim \mathrm{Uniform}(0.8, 1.2)$0, $s \sim \mathrm{Uniform}(0.8, 1.2)$1, $s \sim \mathrm{Uniform}(0.8, 1.2)$2 are computed. Smooth orbit trajectories are parameterized as $s \sim \mathrm{Uniform}(0.8, 1.2)$3, $s \sim \mathrm{Uniform}(0.8, 1.2)$4 with $s \sim \mathrm{Uniform}(0.8, 1.2)$5.

3. Ground-Truth and Instructional Annotations

Each sample is comprehensively annotated to support supervised learning and precise evaluation:

• Spatial ground truth:
  • 3×3 rotation matrix $s \sim \mathrm{Uniform}(0.8, 1.2)$6 and 3×1 translation vector $s \sim \mathrm{Uniform}(0.8, 1.2)$7 for the manipulated object or camera.
  • Object-centric samples: scaling factor $s \sim \mathrm{Uniform}(0.8, 1.2)$8.
  • Camera-centric pairs: relative angles $s \sim \mathrm{Uniform}(0.8, 1.2)$9 and distance $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$0.
• Segmentation and geometry:
  • 2D mask generated by SAM3.
  • 2D bounding box, as $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$1, projected from ground-truth 3D box.
  • Per-pixel depth map $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$2 from the Blender z-buffer, linearized to physical meters.
• Instructions:
  • Textual instruction strings generated automatically, describing the spatial edit (e.g., “Rotate camera 45° to the right and zoom in by 0.5 m”).

4. Dataset Content Statistics

SpatialEdit-500k is balanced across transformation types, object categories, and camera operations, providing broad statistical coverage.

Pipeline	Categories/Scenes	View/Parameter Ranges	Images
Object-Centric	50 object classes	Yaw: {0°, 45°, ..., 315°}; t: ±0.1 m; s: [0.8,1.2]	250,000
Camera-Centric	600 indoor, 400 outdoor scenes	Yaw: {–90°, ... +90°}; Pitch: {–30°, ... +30°}; d: [2,6] m	250,000

Supplementary statistics:

• Object-centric: Each category has ≈5,000 samples; backgrounds: 30% indoor, 40% outdoor natural, 30% outdoor urban.
• Camera-centric: Uniform sampling for each pose parameter; mean aspect ratio of object bounding boxes is ≈1.1, spanning [0.5, 2.0].
• Suggested splits: 450k train / 25k val / 25k test (not explicitly provided).

5. Usage Protocols and Integration for Model Development

Each data instance consists of $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$3, containing the image pair and structured metadata encapsulating all geometric and instructional information.

Principal usage steps:

• Preprocessing:
  • Resize images to $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$4 or the model’s native latent VAE resolution.
  • Normalize pixels to $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$5.
  • Depth and mask channels are optionally loaded for geometry-aware architectures.
  • Instruction strings are encoded with the user’s text encoder (e.g., Qwen3-VL).
• Training and Integration:
  • Object-centric: Randomly select and condition on a transformation (translation, rotation, scaling) from the metadata, input $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$6 to the model, and supervise to predict $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$7.
  • Camera-centric: Input $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$8 and train to predict $$\theta_y \in \{-90^\circ, -45^\circ, 0^\circ, +45^\circ, +90^\circ\}$$9, supervising with both pixel-level loss $$\theta_p \in \{-30^\circ, -15^\circ, 0^\circ, +15^\circ, +30^\circ\}$$0 and pose regression loss $$\theta_p \in \{-30^\circ, -15^\circ, 0^\circ, +15^\circ, +30^\circ\}$$1.
  • Fine-tuning workflows: Apply LoRA or adapter-based specialization by freezing the main backbone and updating only low-rank projection layers.
• Evaluation:
  • Utilize the SpatialEdit-Bench protocol to compute the Moving Score (MS), Rotation Score (RS), Viewpoint Error (VE), and Framing Error (FE), as described in detail in Algorithms 1–2 of the paper (Xiao et al., 6 Apr 2026).

6. Distinctive Contributions and Relevance

SpatialEdit-500k provides the first large-scale, fully synthetic, and precisely annotated dataset expressly designed for benchmarking and training fine-grained spatial image editing models. Its co-development with SpatialEdit-Bench and SpatialEdit-16B establishes a new standard for assessing geometric fidelity and perceptual plausibility in image manipulation. By offering rich 3D supervisory signals and natural-language edit instructions, the dataset enables the training of models capable of resolving both object-centric and camera-centric spatial edits with explicit, measurable accuracy (Xiao et al., 6 Apr 2026).