PaperlayerView exact revision on arXiv

Source-linked AI summary

ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing

Blaž Rolih, Matic Fučka, Filip Wolf, Luka Čehovin Zajc

arXiv:2605.15375v2cs.CVcs.AI

TL;DR

Per-pixel change detection does not model masks as coherent objects, while prior generative methods lag behind discriminative baselines. ChangeFlow generates change masks in latent space using rectified flow and lightweight bi-temporal conditioning. Across four binary benchmarks, it achieves an average F1 of 80.4%, 1.3 points above the previous best.

  • Problem

    Most remote-sensing change-detection methods classify pixels independently, while existing generative methods lag behind discriminative approaches.

    Source: Introduction · p. 1

  • Method

    ChangeFlow generates change masks in latent space via rectified flow, conditioned on simple bi-temporal feature differences from the input images.

    Source: Introduction · p. 2

  • Results

    80.4% average F1 across four binary datasets, 1.3 points above the previous best; on SECOND, ChangeFlow achieves a state-of-the-art F_sc d of 65.9.

    Source: Main results · pp. 5–6

  • Takeaways & Limitations

    Among evaluated methods, ChangeFlow produces masks with the fewest spurious holes and offers a controllable inference speed–accuracy trade-off.

    Source: Conclusion · p. 7, Ablation study · p. 7

    Inference cost scales with the number of ensembled samples.

    Source: Ablation study · p. 7

Abstract

from arXiv · show

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

RSCD localizes changes between images, but change masks are coherent regions rather than independent pixels. ChangeFlow addresses this mismatch by generating masks jointly in a compact latent space.

Introduction

  • RSCD localizes changes between temporal images, but pixel-wise methods treat masks as independent locations, risking fragmented regions and poor global consistency (Rolih et al. 2025; Cheng and Hsu 2025; Yu et al. 2024).Source: Introduction · p. 1
  • Generative models can capture plausible masks and region-level consistency, but prior methods lag discriminative approaches and incur costly pixel-space generation or complex conditioning (Jia et al. 2024).Source: Introduction · pp. 1–2
  • ChangeFlow encodes ground-truth masks with a pretrained VAE, then trains a DiT with rectified flow to transport Gaussian noise to mask latents using simple bi-temporal feature differences.This conditioning avoids auxiliary predictors and elaborate modules.Source: Introduction · p. 2
  • Joint mask generation encourages coherent predictions with fewer holes; adjustable sampling steps enable inference speed-accuracy trade-offs, while multiple samples can be aggregated into a robust prediction.Source: Introduction · p. 2
  • 80.4% average F1 across SYSU, LEVIR, CLCD, and OSCD exceeds ChangeDino by 1.3 points, with ChangeFlow outperforming prior methods on three datasets.Dataset F1 scores are 85.6% on SYSU, 92.1% on LEVIR, 84.5% on CLCD, and 59.5% on OSCD.Source: Introduction · p. 2
  • 65.9 Fscd on SECOND sets a new state of the art, improving over TaCo by 1 point.Source: Introduction · p. 2

Related work

Prior generative change-detection approaches either leave mask prediction to a discriminative network or generate masks through costly pixel-space processes. ChangeFlow instead models whole masks in a compact latent space with rectified flow and lightweight conditioning.

Remote sensing change detection

  • RSCD progressed from pixel-wise differencing and statistical tests to deep architectures, yet pixel-wise changed/unchanged classification remains dominant and can sacrifice robust change-region modelling.The literature includes Siamese convolutional, transformer, state-space, and diffusion-inspired designs.Source: Related work · p. 2

Generative models for computer vision tasks

  • Diffusion and flow models have been applied to Earth Observation and semantic segmentation, but such approaches rarely use the multiple-samples-based inference that generative models offer.Examples include FlowEO for Earth Observation and SemFlow and GSS for semantic segmentation.Source: Related work · p. 2

Generative models for change detection

  • GCD-DDPM and SMDNet (Jia et al. 2024) generate masks in pixel space with costly steps and complex conditioning; ChangeFlow uses latent rectified flow and lightweight feature-difference conditioning.The paper presents this design as enabling more efficient and accurate change-mask generation.Source: Related work · p. 2
  • Other generative change-detection methods synthesize pseudo-changes or extract features from diffusion models, but leave change-mask prediction to a separately trained discriminative network.Pseudo-change generators enlarge and diversify training data; diffusion models trained on remote-sensing imagery serve as feature extractors.Source: Related work · p. 2

Preliminaries

Rectified flow maps Gaussian noise to a target data distribution along a straight-line trajectory. A neural network learns the trajectory’s constant velocity, which is integrated from noise to generate data.

Preliminaries

  • Rectified flow (Liu, Gong, and Liu 2023) maps Gaussian noise X0 to target data X1 along a straight-line trajectory, with intermediate states defined by linear interpolation.X0 ∼ N(0, I), X1 ∼ Pdata, and t ∈ [0, 1].Source: Preliminaries · p. 3
  • A neural network vθ(Xt, t) is trained by mean squared error to predict the trajectory’s constant velocity, X1 − X0.Training samples t from [0, 1].Source: Preliminaries · p. 3
  • At inference, rectified flow generates data by integrating the predicted velocity from noise; its straight path enables accurate few-step integration, supporting its use in modern latent generation (Esser et al. 2024).Source: Preliminaries · p. 3

ChangeFlow

ChangeFlow formulates change detection as latent-space mask generation with rectified flow, guided by bi-temporal features. It extends this approach to semantic change detection and supports sampling-based inference.

ChangeFlow

  • ChangeFlow generates change masks in a compact pretrained-VAE latent space using rectified flow conditioned on shared-encoder bi-temporal feature differences.The absolute-difference conditioning is designed to be lightweight, while joint latent-mask modelling treats each prediction as a coherent object rather than independent per-pixel classifications.Source: ChangeFlow · p. 3, Change Detection as Latent Generative… · p. 4

Change Detection as Latent Generative Synthesis

  • Semantic-mask VAE reconstruction yielded Fscd 99.45 and mIoU 99.53, while binary-mask reconstruction had high F1 and BF1 and low MAE.ChangeFlow learns a joint latent mask velocity field rather than an explicit per-pixel objective.Source: Change Detection as Latent Generative… · pp. 3–4

Inference via Rectified Flow Integration

  • At inference, Euler integration over equally spaced steps generates the mask latent, which the pretrained VAE decoder converts into a single-channel binary mask.The number of integration steps can be adjusted at inference after training.Source: Inference via Rectified Flow Integration · p. 4

Extension to Semantic Change Detection

  • For semantic change detection, ChangeFlow palette-encodes both semantic maps into RGB VAE latents and concatenates them as joint targets.It starts generation from noisy, spatially aligned image latents and conditions on raw bi-temporal features plus the binary change mask.Source: Extension to Semantic Change Detection · pp. 4–5

Sampling-based Ensemble

  • ChangeFlow enables sampling-based ensembling without additional training, using mean aggregation for binary masks and majority voting for semantic masks.The conditional mask distribution is approximated by generating predictions from different initial noise samples.Source: Sampling-based Ensemble · p. 5

Results

ChangeFlow achieves leading binary and semantic change-detection results, while ablations show that its generative objective, ensemble, and lightweight conditioning contribute to performance. Its inference settings trade speed for accuracy, and inference cost scales with ensemble size.

Results

  • Evaluation uses binary F1 on four diverse BCD datasets and mIoU, SeK, and Fscd on SECOND, with metrics calculated from the final-epoch model.Models are evaluated on held-out test sets; BCD inputs are 256 × 256 pixels and SCD inputs are 512 × 512.Source: Results · p. 5

Main results

  • ChangeFlow improves change-region completeness over ChangeDINO and DDPM-CD, while MaskCD’s instance-based predictions remain fragmented across diverse change types.Compared with ChangeDINO it reduces missed detections in homogeneous regions; compared with DDPM-CD it better recovers complete regions and reduces false positives and false negatives.Source: Main results · pp. 5–6, Figure 3 · p. 6
  • 80.4% average F1 puts ChangeFlow 1.3 points above ChangeDINO; it sets state of the art on SYSU, CLCD, and OSCD and runs at 11.8 images per second.Its F1 scores on SYSU, CLCD, and OSCD are 85.6%, 84.5%, and 59.5%; on LEVIR it is within 0.1 points of the best competing method. Methods with higher throughput trail by 3.1–11.3 F1 p.p.Source: Main results · p. 5, Table 2 · p. 6
  • On SECOND, ChangeFlow reaches a state-of-the-art Fscd of 65.9, 1 p.p. above TaCo.Its formulation also yields coherent change regions in the qualitative results.Source: Table 3 · p. 6, Main results · p. 6
  • Across four BCD datasets, ChangeFlow has low structural error, with the fewest spurious holes and highest accuracy in border regions.Structural consistency is measured by error relative to the expected ground-truth number of holes (∆#Holes) and boundary F1.Source: Main results · p. 6, Figure 5 · p. 7

Ablation study

  • At one sample, the generative objective raises average F1 by 0.5 points (79.3 →79.8); five-sample ensembling adds 0.6 points (79.8 →80.4).The single-sample generative model exceeds ChangeDINO by 0.7 points (79.1 →79.8), and fine-tuning the VAE degrades performance, so it remains frozen.Source: Ablation study · pp. 6–7
  • Replacing parameter-free feature difference with complex learnable fusion lowers average F1, and removing either normalization or the absolute value degrades performance.The ablations indicate both conditioning components are essential to the order- and magnitude-invariant signal.Source: Ablation study · p. 7
  • Increasing ensemble repetitions at fixed steps yields more gains than increasing T beyond a small number of steps, while both increase runtime.ChangeFlow uses T = 5 steps and five ensemble predictions; inference cost scales with ensemble size, though one sample already exceeds the previous state of the art.Source: Ablation study · p. 7, Figure 6 · p. 7

Conclusion

ChangeFlow generates change masks in a compact latent space using rectified flow, producing coherent predictions while avoiding pixel-space generation costs. It achieves strong results on binary and semantic change detection benchmarks.

Conclusion

  • 80.4% average F1 across four binary benchmarks, 1.3 points above the previous best, with state-of-the-art results on SYSU, CLCD, and OSCD and competitive results on LEVIR.The four benchmarks are SYSU, LEVIR, CLCD, and OSCD.Source: Conclusion · p. 7
  • 65.9 Fscd on SECOND sets a new state of the art for semantic change detection.Source: Conclusion · p. 7
  • By encoding masks with a pretrained VAE and generating them with rectified flow, ChangeFlow treats each mask as a coherent object while avoiding pixel-space costs and elaborate conditioning.Source: Conclusion · p. 7
  • ChangeFlow produces the fewest spurious holes among evaluated methods and supports a controllable inference speed–accuracy trade-off and sampling-based ensembling without extra training cost.Source: Conclusion · p. 7

Supplementary Material

The supplementary material extends the main manuscript with additional results, analyses, implementation details, and qualitative examples. Its contents span Sections A–E.

Section A

Section B

Section C

  • Section C presents additional qualitative results, including failure cases, semantic change detection results, VAE mask reconstruction, and intermediate-step generation.Source: Supplementary Material · p. 10

Section D

Section E

  • Section E extends model and training implementation details for binary and semantic change detection, and includes ablations, analyses, related methods, and an SCD flow diagram.Source: Supplementary Material · p. 10

A Extended dataset details … Encoder ablations.

Extended dataset details, metric analyses, and ablations broaden the paper’s evaluation; among the reported additional metrics, ChangeFlow has nearly five percentage points higher recall than ChangeDINO and a better precision–recall balance in F1 despite lower precision.

A Extended dataset details; A.1 Data implementation details

  • The benchmarks span varied change scenarios, sensors, GSDs, and dataset sizes, while official train/test splits support reproducibility and fair comparison.Changed pixels are typically under 10%; SYSU and SECOND have higher, but still imbalanced, change ratios. Validation sets for threshold selection are used for SYSU, CLCD, and LEVIR.Source: A Extended dataset details · p. 10, §A.1 · p. 10

B Extended results; B.2 Additional ablations

  • Appendix B extends results and reports additional binary and semantic change detection ablations, with visual results and implementation details referenced in other appendices.The listed binary ablations cover time-step sampling, conditioning resizing, normalisation, conditioning construction, encoder and VAE choices, and binarisation; SCD ablations are in B.3.Source: B Extended results · p. 10, §B.2 · p. 10

B.1 Main results with additional metrics

  • Nearly 5 percentage points higher recall than ChangeDINO; despite lower precision, ChangeFlow achieves a better precision–recall balance in F1.Table 6 reports precision, recall, and F1 across four datasets and their average, with additional precision and recall metrics discussed here.Source: §B.1 · p. 10, Table 6 · p. 11

Training time-step sampling approach

Conditioning resizing

  • Bicubic conditioning resizing performs best overall, although bilinear is better on SYSU; the three tested approaches are relatively similar.The encoder latent is half the VAE latent’s height and width, so the conditioning features must be resized; Lanczos performs worst overall.Source: Conditioning resizing · p. 11, Table 8 · p. 12

Different normalisation layers

  • LayerNorm performs best overall, while BatchNorm and InstanceNorm perform considerably worse for conditioning feature normalisation.The passage attributes this to BatchNorm’s dependence on batch statistics and InstanceNorm’s removal of instance-specific contrast; LayerNorm normalizes channels independently at each spatial location.Source: Different normalisation layers · p. 11, Table 9 · p. 12

Conditioning ablations

  • Absolute-difference conditioning outperforms subtraction and concatenation; concatenation makes differences harder to find, while subtraction is sensitive to temporal order.The ablation also includes a no-normalisation option.Source: Conditioning ablations · p. 12, Table 10 · p. 12

Encoder ablations

  • DINOv3 provides the strongest average features for ChangeFlow; DINOv3 Sat. and DINOv2 perform worse, while RADIO v4 improves on RADIO 2.5 but does not surpass DINOv3.The passage suggests the weaker satellite-pretrained and DINOv2 variants may reflect reduced generalisation from smaller pretraining corpora.Source: Encoder ablations · p. 12

VAE ablations. · Optimal binarisation threshold.

ChangeFlow’s VAE ablations favor the SD-XL VAE with a 4-dimensional latent over tested 16-dimensional alternatives, while a lightweight CNN decoder is generally weaker. For mask binarisation, the validation-set F1 optimum is a two-of-five ensemble rule, with thresholds allowing a precision–recall trade-off.

VAE ablations.

  • The SD-XL VAE with latent dimension 4 achieves the best average performance among the tested pretrained VAEs.Table 12 covers VAEs used for training-label encoding and binary-mask decoding.Source: VAE ablations · p. 12, Table 12 · p. 12
  • The tested 16-dimensional VAEs—SD 3.5, Z-image, and Flux. 1-dev—are consistently slightly worse than SD-XL.The authors suggest higher latent dimensionality may make well-conditioned rectified-flow transport for sparse binary masks harder to learn.Source: VAE ablations · p. 12
  • All tested pretrained VAEs remain competitive overall, supporting off-the-shelf VAE latents as practical representations for change-mask generation.The passage presents this as evidence of their practical effectiveness, despite differences in average performance.Source: VAE ablations · p. 12
  • Replacing the VAE decoder with a lightweight UNet-like CNN decoder is generally weaker on average.The passage does not report numerical differences.Source: VAE ablations · p. 12

Optimal binarisation threshold.

  • On validation data, the optimal F1 binarisation rule marks a region changed when at least two of five ensemble predictions indicate change.Figure 8 shows the two-prediction threshold as the best-performing tested rule.Source: Optimal binarisation threshold · p. 12, Figure 8 · p. 13
  • OSCD and SECOND are excluded from threshold validation because they lack validation sets; the validation-set F1-optimal rule is then used for test evaluation.The passage specifies applying that rule to ChangeFlow’s test-set evaluation in the main paper.Source: Optimal binarisation threshold · pp. 12–13
  • Changing the binarisation threshold lets users prefer recall or precision, while the two-of-five rule provides the best validation-set precision–recall trade-off.Figure 9 reports the trade-off averaged over three datasets.Source: Optimal binarisation threshold · p. 12, Figure 9 · p. 13

B.3 Additional SCD ablations · Colour palette ablations · C Additional qualitative results

The additional sections show how conditioning, latent initialization, and colour encoding choices affect semantic change detection, then catalogue further qualitative comparisons, reconstructions, and generation visualisations.

B.3 Additional SCD ablations

  • Both ΔF and binary mask M are needed for the best SCD results, while M is more compact than ΔF but requires a separate binary-model pass.The mask uses a single channel, compared with 1024 channels for ΔF.Source: §B.3 · p. 13, Table 13 · p. 13
  • Order-invariant conditioning using the absolute value of ΔF and a simpler model without learnable layers outperform removing the absolute value or using a learned 1×1 convolution.Source: §B.3 · p. 13, Table 13 · p. 13
  • Replacing VAE image encodings perturbed by Gaussian noise with pure Gaussian noise substantially worsens SCD performance, but offers no benefit for binary change detection.The authors attribute the binary result to its 2:1 mapping, for which trivial latent choices do not work; they note the SCD insight may extend to other 1:1 dense tasks.Source: §B.3 · p. 13

Colour palette ablations

  • Semantic-mask performance is robust to well-separated colour codes: direct palette reuse matches max-spaced coding on mIoU and is slightly behind on Fscd, while corner coding is on par with direct and slightly below max-spaced.Corner coding assigns each class a vertex of the {−1, 1}³ cube, ensuring pairwise Euclidean distance of at least 2 between class codes.Source: Colour palette ablations · p. 13, Table 14 · p. 13

C Additional qualitative results

C.1 Main qualitative results … D.1 Protocol

The qualitative results show ChangeFlow producing coherent, full change regions, while its VAE mask reconstructions lose little data and coherent regions emerge early in generation. The efficiency appendix reports parameter count, inference time, and GFLOPs under a standardized measurement protocol.

C.1 Main qualitative results

  • ChangeFlow predicts more coherent masks and captures full changed regions with fewer false negatives; no prior method matches this consistently across datasets, consistent with its superior recall.Figure 10 compares the methods across datasets.Source: §C.1 · p. 14

C.2 Additional SCD results

  • Figure 11 shows additional qualitative examples for semantic change detection on SECOND.Source: §C.2 · p. 14

C.3 Failure cases

  • ChangeFlow misses some changed regions, but other methods struggle similarly; in the hardest CLCD example, no model predicts most of the highly semantic changed region.Some SYSU and LEVIR changes may be mislabeled; the misclassified LEVIR case is also shown as uncertain in later confidence visualizations.Source: §C.3 · p. 14

C.4 Visualisation of VAE mask reconstruction

  • An SD-XL pretrained VAE encodes binary masks with minimal data loss and only a minimal reported drop in F1, BF1, and mean absolute error.Figures 13 and 14 visually support mask encoding and RGB-coloring/VAE pass-through with only a few errors; binary images are a special case of RGB images.Source: §C.4 · p. 14

C.5 Intermediate steps visualisation

  • Across five visualized generation steps, a coherent region appears early, while later steps predominantly refine its borders.Source: §C.5 · p. 14

D Computational efficiency

  • The computational-efficiency evaluation reports complementary metrics for ChangeFlow and selected state-of-the-art baselines, including GFLOPs and inference time.The protocol details are in Appendix D.1; additional metrics are in Appendix D.2.Source: D Computational efficiency · p. 14

D.1 Protocol

  • Efficiency is measured by parameter count, inference time (also FPS), and GFLOPs, following (Rolih et al. 2025) with paired 256 × 256 RGB inputs.Measurements use the PyTorch profiler for GFLOPs and an NVIDIA A100-SXM4 40GB GPU with an AMD Epyc 7H12 CPU; inference timing uses 1000 warm-ups and 1000 timed passes, repeated five times, averaging runtime per pass. Models except ChangeFlow use float16 where supported; ChangeFlow uses torch compile because not all modules support float16.Source: §D.1 · p. 14

D.2 Extended computational results

Extended results show that ChangeFlow combines stronger change detection with competitive runtime, while latent, few-step generation avoids the cost of prior pixel-space generation.

D.2 Extended computational results

  • FLOPS can misstate practical speed under parallelisation: ChangeFlow’s 1step5rep setup has higher FLOPS but faster inference than 5step1rep, making FPS a better speed measure.The passage attributes this to multiple repetitions being parallelisable on modern GPUs.Source: §D.2 · p. 14
  • GCD-DDPM is almost 3 orders of magnitude slower than ChangeFlow, using complex auxiliary-model guidance, pixel-space generation, and 1000 steps versus ChangeFlow’s 5 latent steps.Source: §D.2 · p. 14
  • Despite being a true generative method, ChangeFlow runs faster than DDPM-CD and SatDiFuser, whose diffusion networks serve only as feature extractors, while achieving substantially higher change detection accuracy.Source: §D.2 · p. 14
  • ChangeFlow outperforms ChangeDINO by 1.3 points in change detection with comparable throughput and inference time; BTC and HySCDG are faster but perform much worse.Table 15 reports FPS, inference time, parameter count, GFLOPs, and average Precision, Recall, and F1 across four datasets.Source: §D.2 · p. 14, Table 15 · p. 16

E Extended implementation details … E.3 Related methods implementation details

The appendices specify ChangeFlow’s latent-generation components, training and inference procedures, ablation configurations, coherence metrics, and related-method implementations. They describe binary and semantic change detection separately, including the semantic model’s image-latent source and class decoding.

SD-XL VAE.

  • The frozen SD-XL VAE encodes masks in a compact 4-channel latent space, using a 0.13025 scaling factor and RGB-to-single-channel decoding for binary masks.Masks are rescaled to [−1, 1] for encoding and back to [0, 1] after decoding.Source: SD-XL VAE · p. 16

DiT model.

  • The velocity predictor is a LLaMA-2-based DiT with 10 layers, 8 heads, patch size 1, and 1028 input channels; it uses neither class embeddings nor classifier-free guidance.Its output has 4 channels, and its initial learning rate is 1 · 10−4.Source: DiT model · p. 16

Image encoder; Feature difference and normalisation.

  • ChangeFlow uses a finetuned DINOv3 ViT-L encoder and conditions generation on the absolute difference between LayerNorm-normalised bi-temporal features, resized bicubically to the VAE spatial dimensions.The encoder has a 1024-dimensional hidden representation; LayerNorm is applied across channels, and the base model upsamples feature maps by a factor of 2.Source: Image encoder · p. 16, Feature difference and normalisation · pp. 16–17

Ensembling details.

  • Binary inference averages 5 independently sampled predictions after 5-step ODE integration, with the ensemble batched for parallel execution where hardware permits.Each latent is decoded separately before the resulting single-channel masks are averaged.Source: Ensembling details · pp. 17–18

Semantic change detection details.

  • Semantic change detection colourises the land-cover classes with a fixed palette, jointly encodes both timestamps, and conditions generation on binary change guidance and bi-temporal features.At inference, the binary mask is supplied by a separately trained ChangeFlow model and gates the decoded semantic maps; the flow starts from image latents perturbed with Gaussian noise, and class predictions are aggregated by majority vote. In binary experiments, using image latents as the source offered no improvement.Source: Semantic change detection details · p. 18

Other details; E.3 Related methods implementation details.

Encoder ablation details; Conditioning ablation details; Discriminative ChangeFlow ablation details; VAE ablation details.

Coherence analysis details.

  • The coherence analysis measures hole-count deviation and border F1 on binary masks: lower hole deviation indicates greater global consistency, while higher border F1 indicates better border precision.Border F1 is computed on a 3-pixel-wide band around ground-truth masks; the hole procedure excludes border-touching background components and retains enclosed components larger than 10 pixels.Source: Coherence analysis details · p. 20

Related papers

Loading 2605.15375v2…