Research map / Image generation / Text-to-image diffusion
ChangeDiff: A Multi-Temporal Change Detection Data Generator with Flexible Text Prompts via Diffusion Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ChangeDiff, a diffusion-based generator that creates multi-temporal semantic change detection data by generating layouts from multi-class distribution-guided text prompts (T2L) and then images (L2I), with a class distribution refinement loss.
Problem
Semantic change detection needs costly pixel-level annotations; existing change data synthesis methods lack flexible control over change events, need extra segmentation data, and target only binary or specific CD tasks.
Approach
Encode layouts as text prompts of (class name, class ratio) pairs, fine-tune a text-to-layout latent diffusion model with an added cross-attention class distribution refinement loss, then use a ControlNet-style layout-to-image model with noise stitching to produce temporally continuous image pairs without paired images or external data.
Key results
- Augmenting SCanNet on 100% SECOND raises SeK from 22.8 to 24.4 and F1 from 73.1 to 74.1.
- On 5% Landsat-SCD with BiSRNet, SeK rises from 26.1 to 32.4 (as stated in text; table reports 29.4 for BiSRNet+Ours).
- Zero-shot transfer LoveDA to SECOND: ChangeDiff reaches mIoU5 60.1 and SeK5 13.6 versus 53.1 and 7.9 for Changen.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ChangeDiff-generated data | SCanNet · SECOND (5% train) | Sek: 11.3 -> 13.6 | Helped |
| Add ChangeDiff-generated data | SCanNet · SECOND (100% train) | F1: 73.1 -> 74.1 | Helped |
| Add ChangeDiff-generated data | SCanNet · Landsat-SCD (5% train) | Sek: 30.5 -> 33.2 | Helped |
| Add ChangeDiff-generated data | SCanNet · Landsat-SCD (100% train) | F1: 90.6 -> 90.1 | Hurt |
| Add ChangeDiff-generated data | SCanNet · HRSCD (5% train) | F1: 35.7 -> 36.7 | Helped |
| Add ChangeDiff-generated data | SCanNet · HRSCD (100% train) | Sek: 25.1 -> 26.7 | Helped |
| Pre-train on ChangeDiff synthetic data from LoveDA instead of Copy-Paste data | SCanNet (R-18) · SECOND test (zero-shot transfer) | SeK_5: 4.7 -> 13.6 | Helped |
| Replace MCDG-TP with Copy-Paste (ablation) | T2L + L2I · SECOND (5% train; augmentation ablation) | SeK: 13.6 -> 10.6 | Hurt |
| Replace MCDG-TP with original T2L prompts (ablation) | T2L + L2I · SECOND (5% train; augmentation ablation) | SeK: 13.6 -> 11.1 | Hurt |
| Use MCDG-TP instead of Copy-Paste or original T2L prompts (ablation) | T2L + L2I · SECOND (5% train; augmentation ablation) | F1: 64.0 -> 65.3 | Helped |
| Use MCDG-TP (ablation) | T2L + L2I · Landsat-SCD (5% train; augmentation ablation) | F1: 45.2 -> 46.5 | Helped |
| Use MCDG-TP (ablation) | T2L + L2I · SECOND (100% train; fine-tuning transfer ablation) | SeK: 22.8 -> 24.9 | Helped |
Surprising findings
- Changen's binary-change pre-training performs poorly on multiclass tasks despite using 90k pairs, versus ChangeDiff's 10k multi-class pairs.
- Copy-paste based augmentation (ChangeStar, Self-Pair) and GAN-based IAug degrade performance relative to baseline on SECOND.
Limitations
- Relies on sparse-labeled SCD data to fine-tune the T2L model, so each target dataset requires its own fine-tuning.
- Gains on some metrics are small or mixed (e.g., OA and Score decline for some models in tables).
- Evaluated only on semantic change detection with building and land-cover-style classes in remote sensing.
Experimental setup
Datasets: SECOND Landsat-SCD HRSCD LoveDA
Benchmarks: SECOND Landsat-SCD HRSCD
Metrics: SeK Score mIoU OA Kappa F_scd IoU F1 Recall Precision
Baselines: IAug ChangeStar Self-Pair Changen ControlNet + Copy-Paste Copy-Paste ImageNet pre-training
Builds on: latent diffusion models ControlNet Changen Stable Diffusion
Closest papers by meaning
- The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf (CVPR 2025 · similarity 0.87)
- Change3D: Revisiting Change Detection and Captioning from A Video Modeling Perspective (CVPR 2025 · similarity 0.85)
- UniChange: Unifying Change Detection with Multimodal Large Language Model (CVPR 2026 · similarity 0.85)
- Segment Any Change (NeurIPS 2024 · similarity 0.84)
- FM-ChangeNet: Learning Change through Pathwise Feature Transport (NeurIPS 2026 · similarity 0.84)
- DreamDistribution: Learning Prompt Distribution for Diverse In-distribution Generation (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.