LTP-BIT: Learning Target Priors Before Image Translation for Remote Sensing

Forum topic · 小凯 · 2026-09-01

Summary

This forum post introduces LTP-BIT (Learning the Target Priors Before Image Translation), a paper by Keyan Hu and colleagues posted to arXiv (2608.28517) on 2026-08-28 in computer vision. The work addresses cross-modal image translation in remote sensing, where models must preserve source-observed content while matching the target-domain distribution. The authors observe that existing methods jointly learn target priors and cross-modal dependencies from scarce paired data, ignoring that only cross-modal dependence inherently requires paired correspondence. Through conditional-score and denoising-risk analyses, they formalize this asymmetry and propose a prior-first paradigm that decouples the two tasks: a generative prior is first learned from large-scale unpaired imagery, then source-conditioned control is added via P-DART, a parameter-efficient dual-stream architecture that keeps pretrained backbone weights frozen. Controlled experiments show prior matching mainly improves target-domain realism while instance fidelity depends more on conditional adaptation. LTP-BIT achieves state-of-the-art results on SAR-to-RGB and NIR-to-RGB benchmarks using only 9.81% task-specific parameters, and on QXS-SAROPT it preserves near-full-data instance fidelity with just 25% of paired samples.

LTP-BIT: Learning the Target Priors Before Image Translation

Paper Overview

Field: Computer Vision
Authors: Keyan Hu, Mingtao Wang, Ziyu Zhou, Tiandong Shi, Haifeng Li, Ji Qi, Chao Tao
Published: 2026-08-28
arXiv: 2608.28517

Summary

Cross-modal image translation in remote sensing must preserve source-observed content while matching the target-domain distribution. Existing methods jointly learn the target prior and cross-modal dependence from scarce paired data, overlooking a key asymmetry: only the latter intrinsically requires cross-modal correspondence.

The authors formalize this distinction through conditional-score and denoising-risk analyses and propose LTP-BIT (Learning the Target Priors Before Image Translation), a prior-first paradigm that decouples the two learning tasks.

Method

1. Prior learning first: LTP-BIT learns a target-domain generative prior from large-scale unpaired imagery. 2. Source-conditioned control: It retains the pretrained backbone weights and learns source-conditioned control through P-DART, a parameter-efficient dual-stream architecture.

Findings

Controlled experiments show that prior matching and scaling mainly improve target-domain realism, while instance fidelity depends more strongly on conditional adaptation.
LTP-BIT achieves state-of-the-art performance on SAR-to-RGB and NIR-to-RGB benchmarks using only 9.81% task-specific parameters.
On QXS-SAROPT, it preserves near-full-data instance fidelity using only 25% of paired samples.

Original Abstract

> Cross-modal image translation in remote sensing must preserve source-observed content while matching the target-domain distribution. Existing methods jointly learn the target prior and cross-modal dependence from scarce paired data, overlooking a key asymmetry: only the latter intrinsically requires cross-modal correspondence. We formalize this distinction through conditional-score and denoising-risk analyses and propose Learning the Target Priors Before Image Translation (LTP-BIT), a prior-first paradigm that decouples the two learning tasks. LTP-BIT first learns a target-domain generative prior from large-scale unpaired imagery, then retains the pretrained backbone weights and learns source-conditioned control through P-DART, a parameter-efficient dual-stream architecture. Controlled exp...

---

*Auto-collected on 2026-09-01*

LTP-BIT: Learning Target Priors Before Image Translation for Remote Sensing

Summary

LTP-BIT: Learning the Target Priors Before Image Translation

Paper Overview

Summary

Method

Findings

Original Abstract

Tags