UniFlow-RS: Unified Generative Prediction for Heterogeneous Remote-Sensing Tasks

Animated demonstration of task prompts, inputs, ground truth and saved UniFlow-RS predictions across six remote-sensing tasks.

One shared model. Six remote-sensing tasks.
Specify the task with a prompt; predict its structured visual target through the same generative process.

Building extractionChange detectionSemantic segmentationHeight estimationReferring segmentationRotated detection

Abstract

Remote-sensing foundation models increasingly unify representations across sensing modalities and downstream tasks, yet their prediction mechanisms remain largely task-specific. Separate decoders, heads, and objectives leave prediction fragmented across tasks. Can heterogeneous remote-sensing tasks share the prediction process?

UniFlow-RS answers this question with a single latent flow-matching predictor shared across tasks. Heterogeneous outputs are encoded as structured visual targets in a common latent space. Task-dependent conditioning accommodates different input configurations, while a shared Flow Matching objective removes the need for learned task-specific prediction heads and losses. Task–Modality Adaptive Timestep Sampling (TMAS) adapts training timestep allocation to the online learning state of each task–modality group.

A single model handles six tasks across thirteen datasets and remains competitive with task-specific methods. The same checkpoint also transfers across unseen datasets, task settings, occlusion conditions, and near-infrared false-color imagery, without target-specific adaptation.

A shared prediction process

Unify the output space, not just the backbone.

Structured visual targets

Binary masks, class-coded maps, continuous height fields, and rotated-object targets share the same latent prediction space.

Adaptive joint learning

TMAS tracks difficulty and learning activity across task–modality groups to adapt where training samples the flow trajectory.

Deterministic recovery

Task-dependent recovery converts the generated outputs to masks, semantic labels, heights, or rotated instances.

Six tasks, one model

Qualitative results across optical and SAR imagery.

Input, ground-truth and UniFlow-RS predictions for six remote-sensing task families. — Representative qualitative examples from the paper. Selected examples illustrate behavior rather than average performance. Click any result figure to inspect it at full resolution.

Joint adaptation across task families

Task-level scores from the paper. Higher is better in every column.
Setting	Building IoU	Change IoU	Semantic mIoU	Height δ₁	Referring gIoU	Detection mAP50	UTS ↑
Single-task	75.63	73.91	49.26	72.13	72.54	61.39	67.48
Joint	76.38	74.75	51.12	72.06	73.61	62.80	68.45
Joint + TMAS	76.93	74.63	52.57	73.31	73.93	64.75	69.35

Single-task uses one LoRA per task; joint variants share one LoRA. UTS averages the eligible native metrics within each task and then across tasks. See the paper for the dataset-level protocols.

Beyond the training setting

The same checkpoint, without target-specific adaptation.

Language-guided cross-dataset transfer: queries, input images and qualitative comparisons. — Language-guided transfer examples from the paper, with the original queries and reference labels.

From image texture to structured output

A look inside the transport trajectory.

Decoded transport states evolving from the source image into task outputs across six tasks. — Decoded transport states along one inference trajectory per sample. This visualization uses 10 steps to expose intermediate states; the default evaluation uses 2-step inference.

BibTeX

@misc{he2026uniflowrs,
  title={UniFlow-RS: Unified Generative Prediction for
         Heterogeneous Remote-Sensing Tasks},
  author={He, Jianlong and Liao, Liang and Xiao, Jing and
          Chen, Chaofeng and Xia, Gui-Song},
  year={2026},
  eprint={2611.14281},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2611.14281}
}

arXiv:2611.14281 [cs.CV].