CRISP: Progressive Illumination Curriculum for Recurrent Low-Light Video Enhancement with Semantic Priors

Ruirui Lin; Xin Yan Lim; Yini Li; Zhiming Liu; Nantheera Anantrasirichai

cs.CVOct 16, 2026

CRISP: Progressive Illumination Curriculum for Recurrent Low-Light Video Enhancement with Semantic Priors

Authors: Ruirui Lin, Xin Yan Lim, Yini Li, Zhiming Liu, Nantheera Anantrasirichai

Organizations: University of Bristol

Abstract

Low-light video enhancement requires the recovery of brightness, color, and fine spatial structures while preserving temporal coherence in the presence of noise and motion. Existing methods often struggle to balance these objectives, particularly in severely degraded regions. This paper presents CRISP, a recurrent framework trained using a progressive illumination curriculum. Its two-frame recurrent architecture propagates information across successive frames to exploit temporal context beyond the immediate input pair. During training, supervision progressively transitions from intermediate targets interpolated between the low-light input and the reference to the full reference, gradually increasing the required brightness and color correction. This curriculum stabilizes recurrent training without introducing additional inference cost and yields the largest performance improvement among the components evaluated in the ablation study. To further support structural reconstruction, the framework optionally incorporates frozen DINOv3 semantic features through adaptive layer normalization (AdaLN)-style modulation, without fine-tuning the foundation model. Experiments on DID and BVI-RLV demonstrate that CRISP achieves the highest PSNR on DID and the highest SSIM on BVI-RLV among the evaluated methods, while maintaining computational efficiency.

Figures & tables

Figure 1: Overview of CRISP . (a) The recurrent enhancement pipeline combines framewise spatial encoding, multi-scale temporal decoding, and frozen DINOv3 semantic guidance. (b) Expanded view of the semantic-guided spatio-temporal decoder block, incorporating feature warping, CTSF attention, and DINOv3-guided AdaLN modulation.

Stage	Objective	Duration	Supervision target	Active losses
1	Gradual restoration	20 epochs	$Y_{t}^{(\gamma)}$ , $\gamma:0.3\rightarrow 0.8$	Charbonnier + wavelet
2	Brightness and color recovery	30 epochs	$Y_{t}^{(\gamma)}$ , $\gamma:0.8\rightarrow 1.0$	MSE + chroma
3	Full-reference refinement	50 epochs	$Y_{t}$	Charbonnier + wavelet + chroma
4	Temporal consistency	30 epochs	$Y_{t}$	Stage 3 losses + temporal

Table 1: Overview of the four-stage training curriculum. The intermediate target is

Y_{t}^{(\gamma)}=I_{t}+\gamma(Y_{t}-I_{t})

. The curriculum changes training supervision only; the inference architecture remains unchanged.

Method	DID ∗			BVI-RLV			Average
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
SMOID-net [ 52 ]	21.71	0.880	0.194	21.82	0.780	0.280	21.76	0.830	0.237
SDSD-net [ 53 ]	21.88	0.834	0.216	20.60	0.658	0.148	21.24	0.746	0.182
PCDUNet [ 4 ]	18.91	0.725	0.305	24.51	0.845	0.192	21.71	0.785	0.249
STA-SUNet [ 26 ]	19.61	0.714	0.385	26.14	0.873	0.199	22.88	0.794	0.292
TempRetinex [ 13 ]	19.27	0.849	0.314	30.27	0.877	0.327	24.77	0.863	0.321

Table 2: Quantitative comparison (PSNR

\uparrow

, SSIM

\uparrow

and LPIPS

\downarrow

) on the DID and BVI-RLV test sets. Best, second-best, and third-best results are highlighted in red, orange, and yellow, respectively. ∗ DID results may differ from previous works, as we use the provided HD videos, while some works use the 4K versions. † Results reported in the original publication

Method	BVI-Mamba	ReMATF	LLVE-STCD	Ours (Stage3)	Ours (Stage4)
tLPIPS $\downarrow$	0.1604	0.1478	0.0336	0.1039	0.0035
$E_{warp}\downarrow$	0.0319	0.0345	0.0349	0.0342	0.0207

Table 3: Temporal consistency comparison on BVI-RLV.

Figure 2: Spatial quality comparison between the results from Stage 3 (left) and Stage 4 (right) on the BVI-RLV dataset.

Figure 3: Qualitative comparison of the top four methods on the OLiVES dataset, as reported in Table 4 . Clockwise from the top left: BVI-CDM, BVI-Mamba, LLVE-STCD, and ours.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	tLPIPS $\downarrow$	E warp $\downarrow$
SDSD-Net [ 53 ]	22.59	0.716	0.322	0.0600	0.0758
StableLLVE [ 3 ]	22.60	0.850	0.300	0.0280	0.0448
DP3DF [ 27 ]	23.22	0.781	0.325	0.0622	0.0657
BVI-CDM [ 30 ]	26.40	0.854	0.326	0.0167	0.0390
BVI-Mamba [ 31 ]	25.66	0.884	0.228	0.0192	0.0397
LLVE-STCD [ 28 ]	27.00	0.827	0.255	0.0575	0.0534

Table 4: Performance comparison on the OLiVES dataset. Baseline results were provided by the benchmark authors, while CRISP was evaluated independently. All evaluation settings, including the data split, resolution, and metric implementation, were matched.

Panel of Figure 4 — Figure 4: Qualitative comparison on the DID (top row) and BVI-RLV (bottom row) datasets .

Figure 5: Additional qualitative comparison on the DID dataset. Clockwise from the top left: BVI-CDM, DWTA-Net, BVI-Mamba, and our CRISP method.

Panel of Figure 6 — Figure 6: Visual ablation of DINO guidance on the BVI-RLV dataset. The highlighted crop compares the model without DINO guidance against the DINO-guided model and ground truth.

Panel of Figure 7 — Figure 7: Qualitative comparison between representative curriculum checkpoints and the model trained without curriculum learning on the BVI-RLV dataset. Stage 1, Stage 2 and Stage 3 are gradual restoration, brightness and color recovery, and full-reference refinement, respectively.

Configuration	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
w/o DINO guidance	32.7499	0.9154	0.1075
channel-wise $\rightarrow$ spatial-wise attentions	32.7339	0.9134	0.1066
w/o Chroma loss	32.6021	0.9188	0.0961
w/o Curriculum training	31.4143	0.8971	0.1240
Full model	32.7768	0.9229	0.0984

Table 5: Ablation study on BVI-RLV dataset.

	512 $\times$ 512		1920 $\times$ 1080
Architecture	Params (M) $\downarrow$	GMACs $\downarrow$	MEM $\downarrow$	FPS $\uparrow$
BVI-Mamba [ 31 ]	10.00	175.31	-	-
ReMATF [ 54 ]	4.13	164.82	-	-
Ours w/o DINOv3	2.92	121.88	6.51 GB	1.13
Ours Full Model	33.91	194.49	8.70 GB	0.90

Table 6: Complexity analysis at

512\times 512

and

1920\times 1080

Video Restoration

Video restoration reconstructs clearer, more faithful, temporally consistent footage from blur, noise, compression, weather, and other degradations.

See all 59 papers

Dynamic Interpolation for Training-Time Prediction Reshaping in Low-Light Image Enhancement
Oct 13, 2026Peirui Zhu, Shijie Hao, Jingxi Liao +2Low-Light Image Enhancement Image Enhancement
School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China.
VCR: Variance-Driven Channel Recalibration for Robust Low-Light Enhancement
Mar 10, 2026Zhixin Cheng, Fangwen Zhang, Xiaotian Yin +2Image Enhancement Low-Light Image Enhancement
School of Information Science and Technology, University of Science and Technology of China, Hefei 230027, China · Institute of Advanced Technology, University of Science and Technology of China, Hefei 230027, China · CHN Energy Digital Intelligence Technology Development (Beijing) Co., LTD., Beijing, China
Diffusion-Based Low-Light Image Enhancement with Color and Luminance Priors
Feb 27, 2026Xuanshuo Fu, Lei Kang, Javier Vazquez-CorralLow-Light Image Enhancement Conditional Diffusion Models
Computer Vision Center, Cerdanyola del Vallès, Spain and Universitat Autònoma de Barcelona, Cerdanyola del Vallès, Spain

Oct 13, 2026J/K move · Enter open · S save

Dynamic Interpolation for Training-Time Prediction Reshaping in Low-Light Image Enhancement

Peirui Zhu, Shijie Hao, Jingxi Liao, Richang Hong+1

School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China.

Low-Light Image Enhancement Image Enhancement Image Restoration

Open paper AskPDF arXiv

Low-light image enhancement (LLIE) is an important image/video processing problem for reliable visual systems under poor illumination, yet it remains challenging due to the strong coupling among insufficient illumination, noise amplification, and structural degradation. Existing methods typically rely on sophisticated network architectures or engineered objective functions. While effective, the former increases model complexity and deployment burden, whereas the latter relies on empirically tuned loss weights. We revisit supervised LLIE from a training-dynamics perspective and observe that brightness-related errors tend to converge early, while residual noise and structural artifacts often remain insufficiently optimized in later stages. To address this issue, we propose a Dynamic Interpolation Training-only Module (DITM), which introduces no additional learnable parameters and adaptively reshapes the model prediction for supervised feed-forward LLIE training. Specifically, DITM employs brightness-aware interpolation, in which the model prediction is combined with a mean brightness-aligned image constructed by rescaling the model prediction according to the mean brightness of the ground-truth image. By reducing the relative dominance of global brightness discrepancy, DITM makes residual local errors more influential during backpropagation after global luminance has largely stabilized, thereby facilitating artifact suppression and fine-detail restoration in later training stages. DITM is used only during training and is fully removed at inference, introducing no additional inference cost. Extensive experiments on multiple benchmark datasets and CNN-, Transformer-, and Mamba-based LLIE backbones show that DITM consistently improves fidelity and perceptual quality while preserving the inference-time architecture and introducing no additional inference-time computation.

Abstract

Video Restoration

Explore similar work

Dynamic Interpolation for Training-Time Prediction Reshaping in Low-Light Image Enhancement

VCR: Variance-Driven Channel Recalibration for Robust Low-Light Enhancement

Diffusion-Based Low-Light Image Enhancement with Color and Luminance Priors