Research map / Vision-language and multimodal / Multimodal LLMs

UniChange: Unifying Change Detection with Multimodal Large Language Model

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 1 citations (OpenAlex)

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 UniChange, an MLLM-based unified change detection model that uses three special tokens ([T1], [T2], [CHANGE]) to handle both binary and semantic change detection and jointly train on multi-source datasets, reporting SOTA IoU on four benchmarks.

Problem

Existing change detection models are specialised for either binary (BCD) or semantic (SCD) change detection and are trained on a single dataset type, so they cannot jointly exploit multi-source datasets whose class definitions conflict.

Approach

Use an MLLM (LLaVA-7B) to generate special tokens ([T1], [T2], [CHANGE]) whose last-layer embeddings, projected by an MLP, act as instruction-conditioned queries for a SAM-style token-driven decoder that outputs masks; class identity comes from text prompts rather than predefined classification heads.

Key results

• WHU-CD IoU 90.41 and F1 94.96, above ChangeCLIP (90.08 IoU) and other baselines.
• LEVIR-CD+ IoU 78.87 and F1 88.19, versus SFCD-Net 76.12 IoU.
• SECOND SeK 23.02 and mIoU 72.85, exceeding MambaSCD (mIoU 72.73) and HGINet (SeK 21.83); F_scd 63.50 is the best reported, but the text says UniChange achieves 'SOTA' on F_scd and SeK while the paper's Table lists F_scd 63.50 as highest.

Surprising findings

• Table on SECOND reports MambaSCD F_scd of 54.17, far below other methods, indicating binary and semantic metrics can diverge substantially across methods.

Limitations

• Relies on a single backbone family (LLaVA-7B and RSBuilding-ViT-L), with no comparison across MLLM sizes.
• Performance on S2Looking (IoU 53.04) and SECOND (mIoU 72.85) remains modest relative to WHU-CD.
• Joint-training benefit is shown only on a progressive dataset-addition setup, not against alternative multi-task strategies.

Experimental setup

Datasets: WHU-CD S2Looking LEVIR-CD+ SECOND

Benchmarks: WHU-CD S2Looking LEVIR-CD+ SECOND

Metrics: precision recall F1 IoU mIoU F_scd SeK

Baselines: FC-Siam-diff DASNet SNUNet BIT ChangeFormer SAM-CD SCD-SAM SFCD-Net MambaSCD Bi-SRNet

Builds on: LLaVA RSBuilding SAM LoRA

Closest papers by meaning

• The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf (CVPR 2025 · similarity 0.88)
• RSRCC: A Remote Sensing Regional Change Comprehension Benchmark Constructed via Retrieval-Augmented Best-of-𝑁 Ranking (NeurIPS 2026 · similarity 0.87)
• DynamicVL: Benchmarking Multimodal Large Language Models for Dynamic City Understanding (NeurIPS 2025 · similarity 0.85)
• TEOChat: A Large Vision-Language Assistant for Temporal Earth Observation Data (ICLR 2025 · similarity 0.85)
• ChangeDiff: A Multi-Temporal Change Detection Data Generator with Flexible Text Prompts via Diffusion Model (AAAI 2025 · similarity 0.85)
• The STVchrono Dataset: Towards Continuous Change Recognition in Time (CVPR 2024 · similarity 0.84)