ChangeBridge: Spatiotemporal Image Generation with Multimodal Controls for Remote Sensing

Conference: CVPR 2026
arXiv: 2507.04678
Code: https://github.com/zhenghuizhao/ChangeBridge
Area: Remote Sensing / Image Generation
Keywords: Spatiotemporal Image Generation, Diffusion Bridge, Asynchronous Drift, Change Detection, Remote Sensing

TL;DR

Ours proposes ChangeBridge, the first conditional spatiotemporal image generation model for remote sensing. Based on asymmetrically drifting diffusion bridges, it generates post-event images from pre-event images and multimodal conditions (coordinate-text/semantic masks/instance layouts), simultaneously modeling foreground event-driven changes and background temporal evolution, while serving as a data engine for downstream change detection tasks.

Background & Motivation

1. Background: Remote sensing generation methods cover layout-to-image and modality conversion, but conditional spatiotemporal image generation (simulating future scenes based on past observations and multimodal conditions) remains largely unexplored.
2. Limitations of Prior Work: Existing change generation methods only handle event-driven changes (e.g., appearance of new buildings) and cannot model gradual transitions over time (e.g., seasonal changes, vegetation growth).
3. Key Challenge: Two heterogeneous evolutions must be generated simultaneously—intense event-driven changes in the foreground and subtle temporal dynamics in the background. Traditional noise-initialized diffusion models cannot distinguish between the two.
4. Core Idea: (1) Establish a diffusion bridge starting from a composite pre-event state (rather than starting from noise); (2) Assign high drift to the foreground and low drift to the background via a pixel-level drift map (asynchronous diffusion); (3) Employ a drift-aware denoising network.

Method

Overall Architecture

ChangeBridge addresses conditional spatiotemporal image generation in remote sensing: given a pre-event image and multimodal conditions (coordinate-text/semantic masks/instance layouts), it generates a corresponding post-event image. It characterizes two distinct types of changes—foreground intense event-driven changes (e.g., new buildings) and background gradual temporal evolution (e.g., seasonal cycles, vegetation growth). The process does not start from noise but combines the pre-event background and condition-driven foreground as the starting point of the diffusion bridge. Denoising proceeds via a pixel-level drift map allowing fast foreground and slow background changes, finally reconstructing the post-event image through a drift-aware network.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Pre-event Image"] --> C
    COND["Multimodal Conditions<br/>Coord-Text / Semantic Mask / Instance Layout"] --> C
    C["Composite Bridge Initialization<br/>Combine Pre-event Background + Condition Foreground"] --> D
    D["Asynchronous Drift Diffusion<br/>Pixel-level Drift Map: FG γ=1.0 Fast / BG γ≈0.7 Slow"] --> E
    E["Drift-aware Denoising<br/>Drift Map Input for Differentiated Reconstruction"] --> F
    F["Post-event Image"] --> G
    G["Downstream Change Detection<br/>Synthetic Data Engine"]

Key Designs

1. Composite Bridge Initialization: Starting from Pre-event state instead of Noise

Traditional diffusion initializes from pure noise, which destroys background structural information and leads to spatial inconsistency between pre- and post-event states. ChangeBridge combines the pre-event background with the condition-driven foreground as the bridge starting point, ensuring the background spatial structure is preserved throughout the generation process for natural alignment.

2. Asynchronous Drift Diffusion: Multi-speed Fore/Background Evolution

The intensities of foreground event changes and background temporal evolution differ significantly. Uniform drift cannot distinguish between them. The authors assign different drift intensities to each pixel to construct a pixel-level drift map \(\tilde{m}_t(i,j) = m_t \cdot \mathbf{z}_d(i,j)\), using \(\gamma^{fg}=1.0\) for the foreground and \(\gamma^{bg}=0.7\sim0.8\) for the background. This generalizes the Brownian Bridge from uniform drift to spatial asynchronous drift.

3. Drift-aware Denoising: Differential Reconstruction via Drift Maps

Differentiating foreground and background in the forward process is insufficient; the denoising network must also know the target change rate for each pixel. The authors embed the drift map \(\mathbf{z}_d\) into the denoising network to guide differentiated reconstruction, preventing the background from being distorted by intense foreground changes.

4. Multimodal Conditions: Unified Access for Three Control Modes

To support varying granularities of control, the framework integrates three types of conditions: coordinate-text (using rotated bboxes for positioning), semantic masks (color channel mapping for categories), and instance layouts, covering coarse-to-fine multimodal control.

Loss & Training

\[\mathcal{L}_{asy} = \mathbb{E}[\|\tilde{m}_t(\mathbf{z}_a - \mathbf{z}_b) + \sqrt{\delta_t}\boldsymbol{\epsilon} - \boldsymbol{\epsilon}_\theta(\mathbf{z}_t, t, \mathbf{z}_a, \mathbf{z}_c, \mathbf{z}_d)\|^2]\]

The asynchronous drift loss forces the network to predict the transition from pre-event state \(\mathbf{z}_a\) to post-event state \(\mathbf{z}_b\) under the modulation of the drift map \(\mathbf{z}_d\).

Key Experimental Results

Main Results (DiT Variants)

Condition	Dataset	FID↓	IS↑	Spatial Metrics↑
Coord-Text	LEVIR-CC	31.45	5.14	CosSim 0.85
Instance Layout	WHU-CD	40.12	6.77	IoU 78.13
Semantic Mask	SECOND	Best	Best	mIoU Best

Ours outperforms all baselines across all conditions and datasets.

Value as a Data Engine

Training downstream change detection models with synthetic data from ChangeBridge yields significant performance gains, verifying the practical utility of the generated data.

Key Findings

• Asynchronous vs. Uniform Drift: Asynchronous drift significantly improves background temporal consistency.
• Composite Bridge vs. Noise Initialization: Composite bridge preserves spatial structure, enhancing spatiotemporal consistency.
• UNet vs. DiT Variants: DiT variants consistently outperform UNet across all metrics.

Highlights & Insights

• First combination of Diffusion Bridge and Asynchronous Drift: Generalizes Brownian Bridge diffusion to pixel-level asynchronous drift, perfectly matching the design of foreground-fast/background-slow spatiotemporal evolution in remote sensing.
• Validation as Synthetic Data Engine: Demonstrates that ChangeBridge can alleviate the scarcity of training data in change detection tasks.
• Multimodal Condition Framework: Unified support for coordinate-text, semantic masks, and instance layouts.

Limitations & Future Work

• The parameters \(\gamma^{fg}/\gamma^{bg}\) require manual per-dataset tuning.
• Currently only validated in remote sensing scenarios; generalization to natural scenes like street views remains to be explored.
• Spatial resolution of generated images is limited by the reconstruction accuracy of the VQGAN.

Rating

• Novelty: ⭐⭐⭐⭐⭐ The mathematical framework of the asynchronous drift diffusion bridge is elegant with clear physical intuition.
• Experimental Thoroughness: ⭐⭐⭐⭐ Validated on 4 datasets, 3 condition types, UNet/DiT variants, and downstream tasks.
• Writing Quality: ⭐⭐⭐⭐⭐ Complete mathematical derivation and clear illustrations.
• Value: ⭐⭐⭐⭐⭐ Highly significant for remote sensing spatiotemporal simulation and data augmentation for change detection.

Related Papers

• [CVPR 2026] OpenDPR: Open-Vocabulary Change Detection via Vision-Centric Diffusion-Guided Prototype Retrieval for Remote Sensing Imagery
• [CVPR 2026] Spatiotemporal Pyramid Flow Matching for Climate Emulation
• [CVPR 2026] DreamOmni2: Multimodal Instruction-based Generation and Editing
• [CVPR 2026] OmniGen2: Towards Instruction-Aligned Multimodal Generation
• [CVPR 2026] Proxy-Tuning: Tailoring Multimodal Autoregressive Models for Subject-Driven Image Generation