ReFlowSET: Representation-Aligned Latent Flow Matching for SAR-to-EO Image Translation
Jeonghyeok Do, Seungchul Lee, Munchurl Kim
KAIST · Stellarvision Inc.
Published September 1, 2026 · on the feed September 2, 2026 · ▲ 5 · arXiv 2609.00968
The abstract
SAR-to-EO image translation aims to generate electro-optical (EO) imagery from synthetic aperture radar (SAR) observations. Existing latent diffusion approaches typically inherit a predetermined autoencoder, although reconstruction fidelity can vary substantially across codecs and modalities. Because the latent codec affects the round-trip preservation of both SAR conditions and EO targets, codec selection constitutes a fundamental design choice; nevertheless, existing methods largely rely on codecs pretrained on natural images. To remedy this, we introduce ReFlowSET, a conditional latent flow-matching framework that selects its codec through a joint SAR--EO reconstruction audit. Rather than inheriting a heavyweight pretrained generator, ReFlowSET trains a substantially smaller conditional DiT from scratch in the selected latent space, using dual-stream SAR conditioning followed by joint feature refinement. To provide semantic guidance for this from-scratch training, intermediate noisy-EO features are aligned with clean target-EO representations extracted by a frozen vision foundation model. This alignment is used only during training and introduces no additional inference cost. Experiments on QXS-SAROPT and SAR2Opt demonstrate state-of-the-art performance across diverse perceptual fidelity and distributional metrics. Code and pretrained weights are publicly available at https://github.com/KAIST-VICLab/ReFlowSET.
The Brief
2:19A short podcast summary of the paper. Written and voiced by AI.
Transcript
ReFlowSET. Representation-Aligned Latent Flow Matching for SAR-to-EO Image Translation. By Jeonghyeok Do and two others, from KAIST and Stellarvision. Published on the first of September. Its code repository is worked out: about forty source files, with a readme and a licence.
SAR-to-EO translation takes synthetic aperture radar observations and generates electro-optical imagery from them. Latent diffusion approaches to this task typically inherit an autoencoder chosen in advance. Yet how faithfully a codec reconstructs its input can vary a great deal between codecs and between modalities. The codec decides how well both the radar condition and the optical target survive a round trip through the latent space, which makes choosing it a fundamental design decision. Existing methods mostly rely on codecs pretrained on natural images. Radar data is not a natural image, so a codec that serves photographs well may lose exactly what the translation needs, on either side of it.
ReFlowSET is a conditional latent flow-matching framework that treats the codec as something to select, not inherit. It picks its codec through a joint reconstruction audit over both SAR and EO data. The latent space is chosen by how well it preserves both modalities, before any generator is trained in it.
Instead of inheriting a heavyweight pretrained generator, ReFlowSET trains a much smaller conditional DiT from scratch, in the latent space it selected. The radar input enters through dual-stream conditioning, followed by joint feature refinement. Training from scratch needs semantic guidance, and that comes from alignment: intermediate features of the noisy optical image are aligned with clean representations of the target optical image, extracted by a frozen vision foundation model. The alignment is used only during training, so it adds no cost at inference.
On QXS-SAROPT and SAR2Opt, the reported results are state of the art across a range of perceptual fidelity and distributional metrics. Code and pretrained weights are public. Those results come from a generator substantially smaller than the pretrained ones it declines to inherit.
Two habits are set aside at once here: inheriting a codec built for natural images, and inheriting a large pretrained generator. In their place come a codec chosen by measurement and a small model trained from nothing, with a frozen foundation model lending its representations only while training runs.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a codec built for photos, asked to carry radar.
You are generating optical images from radar scans. Like most latent diffusion work, you would encode both into a latent space with an autoencoder pretrained on natural photos.
The road ahead
- 1 · The problem ~7 min Before generating optical images from radar in a latent space, the autoencoder's round trip limits how good any output can ever be.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Choose the codec by how well it rebuilds both radar and optical images, train a small generator from scratch there, and guide it with frozen-vision features.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Pick the codec by audit, encode radar and optical images, process radar in its own stream, then merge and refine both together.
- 6 · The experiment ~6 min You’ll guess: Against prior methods on QXS-SAROPT and SAR2Opt, how did the small from-scratch generator do on perceptual and distributional metrics?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.