NeurIPS 2026 · Poster

Decoupling Direction and Magnitude:
Language-Steered Flow Matching for
Super-Resolution in the Dark

Jiaxin Gao1, Ziyu Yue2, Yaohua Liu3,*,
Danchen Cui4, Zhihui Zhao4, Zhixun Su2

1 The Hong Kong Polytechnic University2 Dalian University of Technology3 The University of Hong Kong4 Taiyuan University of Technology

* Corresponding author: Yaohua Liu · liuyaohua.918@gmail.com

Background and motivation

Super-resolution in the dark is spatially heterogeneous.

A dark low-resolution image is not uniformly corrupted. Some regions are dominated by underexposure and shadow noise, where visibility and structures are severely suppressed; other regions remain relatively reliable and should stay close to the observation. Applying the same restoration strength everywhere is therefore intrinsically risky: weak updates fail in shadows, while strong updates hallucinate textures or distort reliable details.

DeLang-SR revisits this problem from the perspective of flow dynamics. The predicted velocity determines the direction of restoration, while the update magnitude determines how far each latent location moves along that direction. DeLang-SR combines degradation-language conditions and image-specific visual evidence for direction steering, with a spatial update-strength field for magnitude control.

Core method

Direction steering and magnitude control.

DeLang-SR translates measurable low-light statistics into structured degradation language. The language-derived condition is complemented by visual degradation tokens extracted from the input image. Together, they steer the SD3 velocity field, while a degradation-aware spatial field controls the local update magnitude.

Figure 1 from the paper: the complete DeLang-SR framework, with degradation condition construction, visual tokens, and the magnitude-controlled flow update.
Figure 1. Pipeline of DeLang-SR. Degradation language and visual tokens form the fused condition to steer the velocity direction, while the continuous descriptor and local evidence predict the spatial update magnitude.

Degradation conditions from low-light statistics

Measurable statistics cover exposure, contrast, detail, darkness, shadow noise, illumination non-uniformity, and color shift. They are mapped into a continuous descriptor, discrete degradation labels, and a degradation-language prompt:

Equation 2: degradation statistics map to delta, kappa, and the degradation prompt. Delta contains seven entries and the prompt is generated from the discrete labels.

At inference time, degradation conditions are estimated from the low-quality input itself. The language prompt is treated as a coarse restoration intent rather than an oracle annotation. No high-quality reference, manual prompt, or oracle degradation label is required.

Language and visual evidence for direction steering

The degradation prompt provides a global restoration intent. The frozen SD3 text encoder produces the language embedding, while a visual token encoder extracts image-specific evidence from the RGB input, luminance, darkness, local contrast, and global low-light statistics. The two representations are concatenated:

Equation 3: H sub d equals the concatenation of the language embedding h sub ell and visual degradation tokens h sub v.

The fused condition is injected into the SD3 flow transformer to steer the restoration direction.

Spatial update-strength field for magnitude control

The update-strength predictor combines local visual evidence with the continuous degradation descriptor. Its output is bounded and predicted at latent resolution, then broadcast along latent channels:

Equation 4: M phi of y and delta equals sigma min plus sigma max minus sigma min, multiplied by the sigmoid of the visual and descriptor-conditioned predictor.

The field acts as a spatially varying Euler step-size. It controls the local integration magnitude along the predicted restoration velocity:

Equation 1: z hat equals z y minus M phi of y and delta elementwise multiplied by the conditioned velocity; the restored image is obtained by decoding z hat.

Quantitative results

Paired and no-reference evaluation.

We evaluate DeLang-SR on RELLISUR under the ×2 and ×4 settings, measuring fidelity, perceptual quality, exposure recovery, and color consistency. The tables below reproduce the complete comparisons from the paper.

Table 1 from the paper: all-method paired RELLISUR comparisons at x2 and x4, reporting PSNR, SSIM, LPIPS, DISTS, ExpErr, and ColorErr with the original top-three highlighting.
Table 2 from the paper: no-reference RELLISUR comparisons at x2 and x4, reporting NIQE, MUSIQ, MANIQA, and CLIP-IQA with the original highlighting.

Perceptual quality on real DarkFace images

DarkFace evaluates restoration on real nighttime images without paired high-quality references. The paper reports the distribution of four no-reference perceptual quality scores over 100 images.

Figure 4 from the paper: violin distributions of NIQE, MUSIQ, MANIQA, and CLIP-IQA on DarkFace at x2, with per-method means.

Qualitative results

Visual comparisons from the paper.

RELLISUR · ×2 and ×4

The comparison panels below show the original paper's selected regions and method ordering. The top two examples use ×2 super-resolution; the bottom two use ×4.

Figure 5 from the paper: four rows of RELLISUR visual comparisons at x2 and x4 with the paper's original GT region boxes and aligned crops for all methods.

DarkFace · ×2

The original DarkFace panel compares the input, ASID, PFT-SR, CATANet, DHGM, DOD, and DeLang-SR, with the corresponding detail crops.

Figure 6 from the paper: DarkFace x2 visual comparisons and aligned detail crops from the original manuscript.

Citation

BibTeX

@inproceedings{gao2026delangsr,
  title     = {Decoupling Direction and Magnitude:
               Language-Steered Flow Matching for
               Super-Resolution in the Dark},
  author    = {Gao, Jiaxin and Yue, Ziyu and Liu, Yaohua and
               Cui, Danchen and Zhao, Zhihui and Su, Zhixun},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}