ReFlowSET: Representation-Aligned Latent Flow Matching for SAR-to-EO Image Translation

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.

Choose the codec, train from scratch

A SAR–EO reconstruction audit selects the codec; a substantially smaller DiT is trained from scratch in its latent space.

Grid with two QXS-SAROPT rows and two SAR2Opt rows and eight columns: input SAR, SD2.1 FT, BBDM, E3Diff, cBBDM, C-DiffSET, ReFlowSET and ground-truth EO. — **Qualitative comparison on QXS-SAROPT and SAR2Opt.** (a) input SAR, (b)–(f) retrained baselines, (g) ReFlowSET, (h) ground-truth EO.

Same SAR input, every method

Five retrained baselines and ReFlowSET at the native resolution of the paper’s tiles. Select a tile to compare it with the ground truth.

Input SAR ReFlowSET (ours) Ground-truth EO Baseline (retrained)

Use the arrows or the ← → keys to change scene · click a tile to open the comparison sliderSwipe or use the arrows to change scene · tap a tile to open the comparison slider

The codec sets the ceiling

Encode–decode PSNR of six pretrained autoencoders on four SAR–EO datasets: FLUX.2 is best in six of eight dataset–modality settings and raises the mean EO ceiling by 7.49 dB over SD2.1.

Radar chart of encode-decode PSNR for SAR and EO on QXS-SAROPT, SAR2Opt, SpaceNet6 and SAR-1M for the SD2.1, SDXL, SD3.0, SD3.5, FLUX.1 and FLUX.2 autoencoders. The FLUX.2 curve is the outermost in six of the eight settings. — **Modality-wise reconstruction ceilings.** Encode–decode SAR/EO PSNR (dB) per codec and dataset.

Grid with one row per dataset (QXS-SAROPT, SAR2Opt, SpaceNet6, SAR-1M) and seven columns: the reference SAR and EO halves, then the reconstructions of SD2.1, SDXL, SD3.0, SD3.5, FLUX.1 and FLUX.2 with SAR and EO PSNR values in each cell. — **Modality-wise reconstruction ceilings.** Encode–decode SAR/EO PSNR (dB) per codec and dataset.

Within ReFlowSET, replacing the SD2.1 codec with FLUX.2 improves all five metrics on both benchmarks: FID 25.5 → 19.1 on QXS-SAROPT and 84.5 → 66.3 on SAR2Opt (Table 1).

Quantitative results

ReFlowSET achieves the best FID and DISTS on both benchmarks and the best LPIPS on SAR2Opt.

Table 1. Quantitative comparison on QXS-SAROPT and SAR2Opt. All methods are retrained from official code on identical training splits and test pairs. Bold and underline indicate the best and second-best results, respectively.

Bold bestUnderline second best↓ lower is better · ↑ higher is better

Show

Scroll for more columns

Method	Venue	QXS-SAROPT					SAR2Opt
Method	Venue	FID↓	DISTS↓	LPIPS↓	SSIM↑	PSNR↑	FID↓	DISTS↓	LPIPS↓	SSIM↑	PSNR↑
General image translation/restoration methods
pix2pix	CVPR'17	174.6	0.373	0.665	0.203	12.33	261.9	0.347	0.657	0.199	13.39
CycleGAN	ICCV'17	104.4	0.376	0.653	0.262	12.92	143.5	0.330	0.650	0.178	12.90
pix2pixHD	CVPR'18	85.7	0.298	0.573	0.358	16.13	146.3	0.283	0.567	0.268	15.95
SPADE	CVPR'19	90.7	0.292	0.599	0.320	14.53	142.5	0.265	0.597	0.234	14.47
DDPM (SR3)	TPAMI'22	43.8	0.311	0.620	0.359	14.04	122.5	0.295	0.610	0.313	13.65
SD2.1 fine-tune	CVPR'22	19.1	0.257	0.561	0.348	15.40	71.8	0.211	0.541	0.293	16.24
BBDM	CVPR'23	76.6	0.270	0.568	0.352	15.34	143.1	0.290	0.590	0.276	15.29
ControlNet	ICCV'23	50.4	0.307	0.604	0.297	13.42	140.5	0.350	0.643	0.217	11.73
ResShift	NeurIPS'23	140.2	0.334	0.607	0.217	14.20	141.7	0.304	0.597	0.177	14.31
HI-Diff	NeurIPS'23	324.3	0.539	0.692	0.457	17.10	319.8	0.473	0.692	0.384	17.36
StegoGAN	CVPR'24	106.8	0.384	0.658	0.254	12.96	150.1	0.347	0.655	0.158	12.47
SAR-to-EO methods
Cond. Diffusion	GRSL'23	88.6	0.355	0.730	0.213	11.55	211.8	0.415	0.686	0.248	12.48
E3Diff	GRSL'24	47.8	0.278	0.530	0.302	16.44	104.7	0.232	0.529	0.249	16.09
cBBDM	GRSL'25	50.6	0.246	0.539	0.372	16.02	222.3	0.377	0.571	0.361	17.05
C-DiffSET	TCSVT'26	19.9	0.233	0.526	0.380	16.92	78.1	0.214	0.529	0.314	16.81
ReFlowSET (SD2.1)	–	25.5	0.267	0.566	0.337	15.85	84.5	0.217	0.541	0.270	15.82
ReFlowSET (Ours)	–	19.1	0.231	0.534	0.355	16.09	66.3	0.185	0.522	0.287	16.06

On SAR2Opt, ReFlowSET lowers the second-best FID from 71.8 to 66.3 and DISTS from 0.211 to 0.185, relative gains of 7.7% and 12.3%. Checkpoints of the retrained baselines are on Hugging Face.

Ablation and efficiency Table 2

Table 2. Ablation of SAR conditioning, fusion, and EO representation alignment on SAR2Opt. All variants use 12 DiT blocks; D/S denote dual-/single-stream blocks. Params are deployed DiT parameters and exclude training-only alignment modules; VRAM and time report peak training memory and per-image inference latency.

Bold best↓ lower is better

Scroll for more columns

ℒ_Re	SAR Stream	Fusion	Topology	FID↓	Params↓ (M)	VRAM↓ (GB)	Time↓ (sec)
✓	Shared [C-DiffSET]	Input [C-DiffSET]	0D+12S	72.865	108.46	8.18	0.92
✓	Dual	Token	4D+8S	70.450	143.86	13.57	2.06
✗	Dual	Channel	4D+8S	71.851	145.04	9.80	1.34
✓	Dual	Channel	4D+8S	70.436	145.04	10.63	1.34

With alignment enabled, the dedicated SAR stream with channel fusion lowers FID from 72.865 to 70.436; token fusion gives virtually no FID benefit (70.450) at higher VRAM and latency. Adding ℒ_Re lowers FID from 71.851 to 70.436 with unchanged parameters and latency.

Fuse late, align only in training

ReFlowSET framework. Training: the frozen VAE encoder maps the EO and SAR images to latents; a linear trajectory mixes Gaussian noise with the EO latent; a DiT with dual-stream blocks for the SAR latent and the noisy EO latent, followed by concatenation and single-stream blocks, predicts the velocity and is trained with a flow matching loss. Features after the r-th block pass through a projection head and are aligned with EO representations from a frozen foundation model. Inference: an ODE solver integrates the velocity and the VAE decoder produces the EO image. — **ReFlowSET framework.** A conditional DiT learns the latent flow from Gaussian noise to the EO latent under dual-stream SAR conditioning; the alignment branch is removed at inference.

01Codec chosen by reconstruction. Frozen FLUX.2 latents for SAR and EO; a 509.3M-parameter DiT is trained from scratch.
02Dual-stream SAR conditioning. SAR and noisy-EO latents stay separate for 8 of 24 blocks, then fuse channel-wise.
03Training-only alignment. Noisy-EO features are aligned with clean-EO DINOv3 features, at no inference cost.

BibTeX

@article{do2026reflowset,
  title={ReFlowSET: Representation-Aligned Latent Flow Matching for SAR-to-EO Image Translation},
  author={Do, Jeonghyeok and Lee, Seungchul and Kim, Munchurl},
  journal={arXiv preprint arXiv:2609.00968},
  year={2026}
}

Prior work: C-DiffSET (IEEE TCSVT 2026)

@article{do2026cdiffset,
  title={C-diffset: Leveraging latent diffusion for sar-to-eo image translation with confidence-guided reliable object generation},
  author={Do, Jeonghyeok and Lee, Jaehyup and Lee, Seungchul and Kim, Munchurl},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2026},
  publisher={IEEE}
}