Source-linked AI summary
ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing
Blaž Rolih, Matic Fučka, Filip Wolf, Luka Čehovin Zajc
TL;DR
RSCD methods often classify pixels independently, leaving coherent change regions insufficiently modelled and making existing generative approaches costly or complex. ChangeFlow synthesizes masks in a latent space using rectified flow and lightweight bi-temporal conditioning, achieving strong binary and semantic change-detection results. Its generative formulation also supports coherent predictions, adjustable inference, and sample-based ensembling, though inference cost grows with the number of samples.
Problem
Pixel-wise discriminative methods do not directly model change masks as joint structured objects, while prior generative methods lag because pixel-space generation and complex conditioning are costly.
Method
ChangeFlow encodes change masks with a pretrained VAE and uses a rectified-flow diffusion transformer conditioned on simple bi-temporal feature differences to generate mask latents.
Results
80.4% average F1 across four binary datasets and 65.9 Fscd on SECOND establish reported state-of-the-art results, with the binary score 1.3 points above ChangeDINO.
Takeaways & Limitations
The formulation yields spatially coherent predictions, adjustable speed–accuracy at inference, and multiple-mask ensembling without additional training cost.
Takeaways & Limitations
Inference cost scales with the number of ensembled samples, although a single sample already exceeds the previous state of the art.
Abstract
from arXiv · showhide
Remote sensing change detection (RSCD) localises changes between two images of the same geographic region. Most state-of-the-art methods are trained with a per-pixel discriminative objective that classifies each spatial location independently. In this scenario, the predicted changed region is not modelled as a coherent whole, so predictions tend to be spatially fragmented. Generative modelling offers a principled solution: by learning a distribution over plausible change masks, it treats the mask as a single object and encourages global consistency. Yet existing generative RSCD methods lag behind strong discriminative baselines, held back by costly pixel-space generation and overly complex conditioning. We introduce \textbf{ChangeFlow}, which reformulates change detection as the generative synthesis of change masks in a compact latent space via rectified flow, guided by a structured yet lightweight bi-temporal conditioning signal. Changeflow yields spatially coherent predictions without sacrificing efficiency: across four binary benchmarks, SYSU, LEVIR, CLCD, and OSCD, ChangeFlow achieves an average F1 of $80.4\%$, a $1.3$-point gain over the previous best with better efficiency. It also extends to semantic change detection, setting a new state-of-the-art $65.9$ $F_{scd}$ on SECOND. Project page: https://blaz-r.github.io/changeflow_cd
Introduction
Remote sensing change masks have coherent spatial structure, but pixel-wise discriminative methods do not directly model them jointly, producing fragmented predictions. ChangeFlow addresses this gap by generating masks in latent space with rectified flow and lightweight bi-temporal conditioning.
- RSCD localises changes between images of the same geographic region acquired at different times.
- Pixel-wise supervision treats locations independently even though change masks delineate coherent objects and boundaries.
- Generative modelling learns a conditional distribution of plausible masks, enabling direct capture of long-range dependencies and region-level consistency.
- Existing generative methods remain behind discriminative approaches because pixel-space generation is costly and conditioning pipelines are complex.
- ChangeFlow encodes ground-truth masks with a pretrained VAE and trains a diffusion transformer with rectified flow to generate mask latents from Gaussian noise.
- ChangeFlow jointly generates masks, supports adjustable inference steps and multiple-mask sampling, and achieves 80.4% average F1 across four binary datasets.
Related work
RSCD has progressed from pixel-wise and statistical methods to end-to-end deep architectures, including convolutional, transformer, state-space, diffusion-inspired, and pretrained models.
- RSCD evolved from pixel-wise differencing and statistical tests to end-to-end deep models.
- Siamese networks span convolutional and transformer architectures, with later work also using state-space and diffusion-inspired backbones.
- Large-scale pretraining is increasingly important for performance and robustness in RSCD.
Preliminaries
Rectified flow transports Gaussian noise to a target data distribution along a straight-line trajectory. A neural network learns the constant velocity field, which is integrated during inference in few steps.
- Rectified flow maps Gaussian noise X0 ∼ N(0, I) to a target distribution X1 ∼ Pdata via a straight-line trajectory.
- Because the trajectory has constant velocity X1 − X0, the velocity network is trained with mean squared error.
- During inference, integrating the predicted velocity field from noise generates data in few steps.
ChangeFlow
ChangeFlow performs change-mask synthesis in a compact VAE latent space using rectified flow and lightweight bi-temporal feature-difference conditioning. Its inference integrates a velocity field to decode coherent binary or semantic change predictions.
- ChangeFlow moves generative change detection from pixel space to compact latent space, using a pretrained VAE and rectified flow.The VAE encodes masks, while the rectified-flow model transports Gaussian noise to mask latents.
- Bi-temporal conditioning uses the absolute difference between layer-normalised features from a shared-weight encoder, avoiding complex auxiliary conditioning.The difference signal is concatenated with the intermediate latent and supplied to a DiT-based vector-field model.
- The model jointly learns the mask latent rather than an explicit per-pixel objective, treating the prediction as a coherent object.This formulation is intended to capture spatial structure directly rather than factorising the prediction into independent pixel terms.
- At inference, ChangeFlow integrates the rectified-flow ODE with Euler steps, decodes the final latent through the VAE, and averages RGB channels into a binary mask.The number of inference steps can be adjusted after training, providing a flexible sampling schedule.
- Extension to Semantic Change Detection: For semantic change detection, the method encodes colourised semantic maps jointly and conditions generation with raw bi-temporal features plus binary-change guidance.Generated semantic latents are mapped back to class indices after VAE decoding.
- ChangeFlow supports sampling-based inference by generating multiple masks from different initial noises and aggregating them by mean or majority vote.The rectified-flow formulation implicitly defines a conditional distribution over change masks.
Results
ChangeFlow achieves strong binary and semantic change-detection performance while maintaining controllable efficiency and improving structural coherence. Ablations indicate that both the generative objective and inherent ensembling contribute to performance.
- Binary change detection: 80.4% average F1 across four BCD datasets, 1.3 points above ChangeDINO, with state-of-the-art results on SYSU, CLCD, and OSCD.The method reaches 85.6%, 84.5%, and 59.5% F1 on SYSU, CLCD, and OSCD, respectively, and remains within 0.1 points of the best method on LEVIR.
- Efficiency: 11.8 images per second makes ChangeFlow roughly 3 images per second faster than ChangeDINO, while faster methods trail by 3.1–11.3 F1 points.FPS includes all sampling steps and ensembling, benchmarked on an NVIDIA A100.
- Semantic change detection: 65.9 Fscd on SECOND establishes a new state of the art, 1 point above the previous best TaCo.The qualitative results also show good coherence within changed regions.
- Coherence: ChangeFlow yields low structural error, with the fewest spurious holes and highest border-region accuracy across four BCD datasets.Coherence is evaluated using hole-count error and boundary F1.
- Ablations: The generative objective raises average F1 by 0.5 points, while aggregating five samples adds a further 0.6 points.Average F1 changes from 79.3 to 79.8 with the generative objective and from 79.8 to 80.4 with ensembling; fine-tuning the VAE degrades performance.
- Inference trade-offs: Increasing inference repetitions improves performance more than increasing rectified-flow steps beyond a small number, and both increase runtime.This exposes a controllable speed–accuracy trade-off at inference time.
Conclusion
ChangeFlow recasts remote sensing change detection as latent change-mask synthesis using a pretrained VAE and rectified flow. It delivers coherent predictions, adjustable inference trade-offs, and strong results across binary and semantic benchmarks.
- Conclusion: ChangeFlow synthesizes change masks as coherent latent objects rather than classifying pixels independently.Masks are encoded with a pretrained VAE and generated using rectified flow.
- Conclusion: Across four binary benchmarks, ChangeFlow reaches 80.4% average F1, improves on the previous best by 1.3 points, and remains competitive on LEVIR.It sets a new state of the art on SYSU, CLCD, and OSCD.
- Conclusion: ChangeFlow achieves a new state-of-the-art 65.9 Fscd on SECOND for semantic change detection.The formulation therefore extends beyond binary change detection within the reported evaluation scope.
Supplementary Material
The supplementary materials extend the main paper with additional datasets, metrics, ablations, qualitative analyses, efficiency evaluations, and implementation details.
- Supplementary Material: The supplement provides extended dataset details and additional BCD and SCD ablations.It also includes additional precision and recall metrics.
- Supplementary Material: Additional qualitative results include failure cases, more SCD results, VAE mask reconstruction, and intermediate-step generation.These materials extend the qualitative analysis beyond the main manuscript.
- Supplementary Material: The supplement documents computational-efficiency protocols, extended results, and implementation details for binary and semantic settings.It includes a flow diagram for the method and training details.
A Extended dataset details
The appendix details diverse datasets, evaluation protocols, and ablations showing how ChangeFlow’s training, conditioning, and normalization choices affect performance.
- A Extended dataset details: The benchmarks span diverse change scenarios, sensors, ground sampling distances, and dataset scales.They include building and urban expansion changes, cropland-limited changes, and a few hundred to several thousand image pairs.
- A Extended dataset details: Changed pixels typically comprise less than 10% of images, creating severe class imbalance in most datasets.SYSU and SECOND have higher, but still relatively unbalanced, change ratios.
- A Extended dataset details: ChangeFlow uses official train and test splits, with validation sets from SYSU, CLCD, and LEVIR for threshold selection.OSCD and SECOND lack validation sets for this purpose.
- A Extended dataset details: Logit-normal time sampling emphasizes training near timestep 0.5, where uniform sampling performs consistently worse across datasets.The sampling maps N(0, 1) through a sigmoid to concentrate times around the halfway point between noise and data.
- A Extended dataset details: An ensemble predicts a region as changed when at least two of five masks indicate change, yielding the best F1-based binarization strategy.The method also uses bicubic resizing and LayerNorm for conditioning; LayerNorm outperforms InstanceNorm and BatchNorm.
- A Extended dataset details: DINOv3 provides the strongest average encoder features, while absolute-difference conditioning outperforms subtraction, concatenation, and complex alternatives.These findings support the paper’s simpler feature-difference conditioning design.
VAE ablations.
The VAE ablations examine mask encoding choices and binarization, finding strong performance with compact pretrained representations and ensemble thresholding.
- VAE ablations: The SD-XL VAE with latent dimension 4 achieves the best average performance among pretrained VAEs.VAEs with latent dimension 16 are consistently slightly worse, plausibly because higher dimensionality complicates rectified-flow transport for sparse binary masks.
- VAE ablations: A lightweight UNet-like CNN decoder is generally weaker on average than the evaluated pretrained VAE decoders.
- VAE ablations: Binarizing an ensemble of five masks when at least two predictions indicate change gives the optimal validation-set F1 strategy.OSCD is excluded because it has no validation set.
B.3 Additional SCD ablations
Additional semantic change detection ablations evaluate conditioning, initialization, and colour encoding, favoring compact order-invariant signals and well-separated class codes.
- B.3 Additional SCD ablations: Both feature differences and the binary change mask are necessary for the best semantic change detection results.The binary mask is more compact but requires a separate binary-model pass.
- B.3 Additional SCD ablations: Absolute-value feature differences and simpler conditioning outperform order-sensitive or learnable alternatives.Removing the absolute value or adding a learnable 1×1 convolution performs worse.
- B.3 Additional SCD ablations: Gaussian-noise-perturbed image latents outperform pure Gaussian-noise initialization for semantic change detection.The authors relate this to the spatial alignment of image latents with semantic targets and report no analogous benefit in the binary setting.
- B.3 Additional SCD ablations: Semantic-map performance remains robust across colourisation schemes when class codes are well separated.Direct palette reuse and binary corner coding are close to greedy max-spaced RGB values, with slightly lower Fscd.
C.1 Main qualitative results
Qualitative results show coherent generated masks, accurate coverage of changed regions, and early formation of regions followed by boundary refinement, alongside documented failure cases.
- C.1 Main qualitative results: ChangeFlow predicts more coherent masks and captures fuller changed regions with fewer false negatives than prior methods across multiple datasets.The qualitative pattern agrees with its superior recall reported in the appendix.
- C.1 Main qualitative results: ChangeFlow still misses some changed regions, and the hardest CLCD example defeats all compared models on most of the changed area.Some SYSU and LEVIR errors may reflect mislabelling, with uncertainty visible for one LEVIR case.
- C.1 Main qualitative results: RGB-trained SD-XL VAE encoding and decoding introduces minimal information loss for binary change masks.The paper reports minimal drops in F1, BF1, and mean absolute error, supported by visual reconstructions.
- C.1 Main qualitative results: The coherent changed region appears early during five-step generation, while later steps primarily refine borders.
- C.1 Main qualitative results: Efficiency is evaluated using parameter count, inference time or FPS, and GFLOPs.Inference timing uses 256 × 256 RGB input pairs and repeated warm-up and timed passes.
D.2 Extended computational results
ChangeFlow combines stronger change-detection performance with practical inference efficiency. Its latent generation is substantially faster than the previous generative method and can outperform feature-extraction diffusion baselines.
- 1.3 points: ChangeFlow outperforms ChangeDINO in change detection while offering comparable throughput and inference time.BTC and HySCDG infer faster, but perform substantially worse in change detection.
- Almost 3 orders of magnitude: ChangeFlow is faster than GCD-DDPM, using latent generation and 5 steps instead of pixel-space generation and 1000 steps.GCD-DDPM also uses a complex auxiliary conditioning scheme for guidance.
- Substantially higher change detection accuracy: ChangeFlow is faster than DDPM-CD and SatDiFuser despite performing true generative prediction.DDPM-CD and SatDiFuser use generative diffusion networks as feature extractors rather than for direct mask generation.
- FPS, inference time, parameter count, FLOPS, Precision, Recall, and F1 are reported for computational comparison across the evaluated models.The results use an Nvidia A100-SXM4 40GB GPU and the stated evaluation protocol.
E Extended implementation details
The implementation uses a frozen SD-XL VAE, a DiT rectified-flow predictor, and normalized absolute feature differences for conditioning. Supplementary details cover ensemble inference, semantic-mask encoding, decoding, and coherence evaluation.
- Supplementary setup: The supplementary material organizes implementation, ablation, analysis, and related-method execution details across its appendices.Experiments were conducted on an NVIDIA A100-SXM4 40GB GPU and AMD Epyc 7H12 CPU on a Slurm cluster.
- Implementation details: The frozen SD-XL VAE maps masks into a compact 4-channel latent space for generation.Binary masks are repeated across RGB channels for encoding and averaged back to one channel after decoding.
- Implementation details: The DiT uses a 256-channel, 10-layer, 8-head architecture with patch size 1 and no class embeddings or classifier-free guidance.Its input channels combine the image-encoder dimension c with the VAE latent dimension d.
- Implementation details: Conditioning uses LayerNorm-normalized, per-element absolute differences between the two image feature maps.The resulting conditioning vector is resized to the VAE spatial dimensions using bicubic interpolation.
- Implementation details: Five independently sampled noise vectors produce five mask hypotheses that are decoded and averaged into the binary prediction.The repetitions are batched and can be parallelized, reducing the practical overhead of ensembling.
- Semantic change detection: Semantic masks use a fixed RGB palette with white assigned to the no-change class and six SECOND classes represented by remaining colors.The semantic case uses nearest-neighbour decoding and per-pixel majority voting, unlike binary averaging.
- Implementation details: Binary ensemble outputs are thresholded at 0.3, equivalent to requiring at least two of five discretized hypotheses to indicate change.This converts the continuous ensemble prediction into a binary mask.
- Coherence analysis: Structural coherence is evaluated by penalizing holes and border imprecision, with lower deviation and higher BF1 indicating better consistency and border precision.