CLEAR: Color-Texture Latent Alignment for Single Image Reflection Removal

Abstract

Single Image Reflection Removal (SIRR) seeks to recover the transmitted scene from a photograph captured through glass. Recent latent restoration methods use pretrained generative priors, but many recover detail by feeding high-frequency features from the degraded input back into the decoder. This works for degradations such as blur, noise, or haze, where input detail mostly belongs to the scene. It is less suitable for SIRR, where reflections add a second structured layer: directly reusing the input's skip connections can sharpen the image while also bringing reflection edges back. Latent compression also introduces color drift and fine-detail loss.

We propose CLEAR (Color-Texture LatEnt Alignment for Single Image Reflection Removal), a deterministic, single-timestep, non-iterative latent framework that uses a pretrained latent diffusion backbone as a natural-image prior. A physics-guided separator organizes the prediction around the transmission, reflection, and residue formation model and outputs coarse layer latents in one forward pass. The Hybrid Feature Alignment Module (HFAM) then treats input-derived skips as contaminated layer-mixture features. It rectifies their high-frequency content with a clean-transmission-aligned color query, learned from the input features themselves under ground-truth feature supervision, keeping transmission-consistent texture while attenuating reflection-colored detail. CLEAR achieves the best average across fidelity and perceptual metrics on six real-world benchmarks and the lowest LPIPS on every dataset, with the largest gains on the hardest benchmarks.

Method

CLEAR is a two-stage pipeline: a dual-branch latent separator (Stage 1) and a cross-latent decoder augmented with the Hybrid Feature Alignment Module (Stage 2). Both stages run in a single deterministic forward pass — no sampling, no noise schedule.

CLEAR pipeline: Stage 1 latent disentanglement with transmission and reflection branches, Stage 2 feature alignment with HFAM. — **Figure 1 — Pipeline overview.** **Stage 1 (latent disentanglement):** the input is encoded by a frozen SDv2 VAE and processed by two U-Net branches that exchange information via latent cross-attention, predicting transmission and reflection/residue latents in a single deterministic step. **Stage 2 (feature alignment):** the transmission latent is decoded by a cross-latent decoder augmented at every scale by the HFAM, which injects high-frequency detail and chromatic cues. During training the HFAM is supervised with ground-truth features extracted by the same frozen encoder. Snowflakes mark frozen modules; trainable modules are unmarked.

Stage 1 — Physics-guided latent separator

The image formation model I = T + R (+ residue Φ) organizes the prediction. Two SDv2-initialized U-Net branches exchange information through latent cross-attention and output the transmission, reflection, and residue latents in one forward pass. The VAE encoder stays frozen, so both branches share the same latent manifold and the same natural-image prior.

Stage 2 — HFAM decoder refinement

Stage 1 separates layers structurally but leaves blurred texture and color drift. The decoder needs a high-frequency source, and the input is the only adequate one — but the input skips are contaminated layer-mixture features. HFAM turns skip injection into an alignment problem.

HFAM at one decoder scale: color query, structural key, and DWT high-frequency value feeding a color-gated channel attention. — **Figure 2 — HFAM at one decoder scale.** Input skip Fⁱ yields a color query Q, a structural key K, and a high-frequency value V (discrete wavelet transform); color-gated channel attention reweights V toward clean-transmission statistics (supervised during training, computed from Fⁱ at inference). Applied at every decoder scale.

Deterministic. Single timestep, non-iterative, no random sampling — the output is reproducible run to run.
Color-gated channel attention. Affinity is computed across channels (Q̂K̂^⊤ ∈ ℝ^C×C), so cost is linear in the number of pixels.
Clean-transmission supervision. The color query and high-frequency heads are supervised with features of the clean transmission extracted by the same frozen encoder — the targets never touch the decoded output.
Physically grounded. The learned residue Φ absorbs non-linear effects such as saturation, where the linear model I = T + R is violated.

Results

We evaluate on six real-world benchmarks — Nature, SIR2, Real20, DRR-C, DRR-S, and OpenRR-1k — against RobustSIRR, RRW, DSIT, DSRNet, RDNet, DAI, and WindowSeat. Following the adaptation protocol of DAI, every method is fine-tuned on identical training data and evaluated with the same scripts, so the comparison isolates architectural and methodological differences rather than training-data advantages.

Qualitative comparison

Every tile pairs the input (left) with the method's result (right). Click a tile to load it into the slider above, double-click for a full-size comparison, or use ← / → to step through the methods.

Quantitative comparison

↑ higher is better, ↓ lower is better. Bold = best, underlined = second best. Averaged over all six benchmarks (1,040 images), CLEAR is best on every metric and has the lowest LPIPS on every individual dataset.

Dataset	Metric	RobustSIRR	RRW	DSIT	DSRNet	RDNet	DAI	WindowSeat	CLEAR (Ours)

Method	PSNR ↑	SSIM ↑	LPIPS ↓	MANIQA ↑	QUALICLIP ↑
RobustSIRR	21.14	0.791	0.221	0.479	0.704
RRW	23.43	0.815	0.202	0.458	0.689
DSIT	23.99	0.817	0.198	0.479	0.706
DSRNet	21.42	0.809	0.201	0.485	0.714
RDNet	23.92	0.822	0.192	0.486	0.707
DAI	25.98	0.869	0.159	0.460	0.696
WindowSeat	25.48	0.810	0.208	0.465	0.674
CLEAR (Ours)	26.59	0.906	0.106	0.504	0.731

Average over Nature, SIR2, Real20, DRR-C, DRR-S and OpenRR-1k (1,040 images).

Ablations & analysis

Stage 2, residue, and HFAM sub-paths

S2	Φ	HF	Color	Nature		Real20		SIR2
S2	Φ	HF	Color	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑
✗	✗	✗	✗	25.09	0.838	23.52	0.779	26.34	0.904
✓	✗	✗	✗	25.32	0.847	23.52	0.819	26.41	0.932
✓	✓	✗	✗	26.02	0.865	23.60	0.824	26.57	0.933
✓	✓	✗	✓	26.15	0.869	23.52	0.823	26.65	0.934
✓	✓	✓	✗	26.04	0.868	23.73	0.823	26.69	0.932
✓	✓	✓	✓	26.45	0.867	23.74	0.824	27.71	0.934

The color and high-frequency paths are complementary: alone each improves PSNR, together they are best. The full model also reduces ΔE₀₀ by 4.2% over the Stage-2+Φ baseline on SIR2 (4.983 → 4.774) and cuts reflection leakage (corr.) from 0.2657 to 0.2243 (−15.6%, paired bootstrap, 10k resamples).

Capacity vs. methodology

Method	Total	Trainable	Time (s) ↓	PSNR ↑	SSIM ↑	LPIPS ↓
DAI	1.74B	0.95B	0.40	25.98	0.869	0.159
GenSIRR †	12.11B	12.02B	53.8	26.47	0.879	0.130
CLEAR (BK-SDM-Small)	1.29B	1.08B	0.42	26.24	0.908	0.117
CLEAR (SDv2)	2.27B	1.84B	0.56	26.59	0.906	0.106
CLEAR (generic S2 refiner) ‡	2.27B	1.84B	0.57	25.80	0.861	0.152
CLEAR (half-size HFAM) ‡	2.27B	1.82B	0.54	26.25	0.903	0.108

Averaged over all benchmarks; time on one A6000. † GenSIRR is concurrent and uses its native 1MP setting. ‡ A generic Stage-2 refiner matched to HFAM's parameters and FLOPs, and a half-size HFAM, both trained under identical schedules. The gains come from the method, not from scale.

How detail enters Stage 2

Detail injection (same Stage 1 output)	PSNR ↑	SSIM ↑	LPIPS ↓
Stage 1 only (no Stage 2)	25.96	0.891	0.115
Stage 1 + copy input skips	26.21	0.896	0.112
Stage 1 + Wavelet fix	22.54	0.825	0.160
Stage 1 + AdaIN fix	23.42	0.825	0.158
Stage 1 + HFAM (ours)	26.59	0.906	0.106

Copying the input skips recovers sharpness but reintroduces the reflection; global color fixes fall below the no-refinement baseline. HFAM is best on all three metrics.

Citation

If you use this work, please cite:

@inproceedings{vo2027clear,
  title     = {CLEAR: Color-Texture Latent Alignment for Single Image Reflection Removal},
  author    = {Vo, Tu and Zaheer, Sheir A. and Park, Chan Y.},
  booktitle = {The IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}