DIRS: Principled Reflection Separation

Abstract

Nonlinearity Meets Interaction

DIRS tackles real-world layer entanglement by introducing a learnable nonlinear superposition model and unifying dual-stream architectures under a generalized interaction paradigm.

The Challenge: Reflection superimposition remains a severely ill-posed problem. Existing methods often struggle in real-world scenarios because they rely on flawed linear blending assumptions in sRGB space or treat layer disentanglement as isolated, single-stream subproblems.

Nonlinear Formation: We challenge the conventional linear composition model. DIRS introduces a learnable nonlinear interaction term that faithfully captures the complex layer couplings and biases introduced by real-world ISP pipelines.

Unified Interaction: To resolve the intrinsic ambiguity, we propose a generalized dual-stream interactive architecture. It facilitates deep, bidirectional feature exchange between transmission and reflection pathways. This principled framework seamlessly unifies activation-based, gate-based, and attention-based mechanisms across both CNN and Transformer backbones.

One Framework, Three Variants: DIRS provides highly configurable solutions, including an efficient activation-based CNN (YTMT), a mutually gated CNN (MuGI), and a state-of-the-art Transformer with dual-stream joint attention (PAIR).

Method

DIRS Framework

The Learnable Offset-Residual model captures complex nonlinear physical couplings, while the interactive dual-stream architecture enables explicit bidirectional feature exchange to cleanly disentangle the overlapping layers.

DIRS-YTMT

Activation-based CNN interaction that recycles suppressed features between streams.

DIRS-MuGI

Mutually gated CNN interaction for spatially varying nonlinear reflection coupling.

DIRS-PAIR

Transformer interaction with dual-stream joint attention and parallel self-attention.

Models

Model Zoo

FLOPs and latency are measured on a single NVIDIA RTX 3090 with 256 x 256 inputs. PSNR/SSIM are averaged over Real20 and SIR².

3released DIRS variants

26.95DIRS-PAIR + Nature average PSNR

0.926DIRS-PAIR + Nature average SSIM

Model	Type	Params	FLOPs	Time	PSNR	SSIM
DIRS-YTMT	CNN, activation interaction	32.42M	102.91G	31.35 ms	24.94	0.902
DIRS-MuGI	CNN, mutual gating	84.47M	153.98G	49.95 ms	25.63	0.913
DIRS-PAIR	Transformer, joint attention	48.80M	200.22G	75.36 ms	26.37	0.918
DIRS-PAIR + Nature	Transformer, joint attention	48.80M	200.22G	75.36 ms	26.95	0.926

Survey

Interactive Roadmaps

Decades of reflection removal and separation methods organized by input modality, physical cue, prior constraint, and network paradigm.

Single-Image Reflection Removal / Separation

Multiple-Image Reflection Removal / Separation

Results

Visual Results

DIRS separates strong real-world reflections and extends naturally to reflection scene reconstruction and polarized image reflection separation.

Real-world reflection separation comparison — Transmission predictions in challenging real-world scenarios.

Polarized reflection separation — DIRS adapted to polarized multi-image reflection separation.

Examples

Reflection

Reflection Scene

Citation

BibTeX

Please cite DIRS if this project is useful for your research.

@article{hu2026dirs,
  title={Principled Reflection Separation via Nonlinear Superposition and Feature Interaction},
  author={Hu, Qiming and Li, Mingjia and Li, Yuntong and Guo, Xiaojie},
  journal={arXiv preprint},
  year={2026}
}