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

Peirui Zhu; Shijie Hao; Jingxi Liao; Richang Hong; Meng Wang

Fig. 1: Empirical motivation for DITM. (a) Global brightness discrepancy is reduced primarily during the early stage of baseline training. We present the evolution of the global brightness discrepancy throughout the training of four representative baseline LLIE models on LOLv1 and LOLv2-synthetic, evaluated using the Y channel of YCbCr and the V channel of HSV. Let

A_{t}

denote the baseline prediction at training stage

t

, and let

Y

denote the corresponding ground-truth image. For brightness representation

c

, BMAE is defined as

\mathrm{BMAE}_{c}(t)=\left|\mathrm{Mean}\!\left[\Phi_{c}(Y)\right]-\mathrm{Mean}\!\left[\Phi_{c}(A_{t})\right]\right|

, where

\Phi_{c}(\cdot)

extracts the specified brightness component,

\mathrm{Mean}[\cdot]

computes its spatial mean over all pixels, and

c\in\{\mathrm{YCbCr\text{-}Y},\mathrm{HSV\text{-}V}\}

. At each training iteration, BMAE is computed by averaging the brightness discrepancy over the samples used for the current update. The early brightness-error reduction ratio is defined as

\eta_{c}(r)=\frac{\mathrm{BMAE}_{\mathrm{init}}^{c}-\mathrm{BMAE}_{r}^{c}}{\mathrm{BMAE}_{\mathrm{init}}^{c}-\mathrm{BMAE}_{\mathrm{final}}^{c}}\times 100\%

, where the three BMAE terms denote the averaged BMAE values around the initial stage, training progress

r

, and the final stage, respectively. The annotated

\eta_{c}(20\%)

values show that 72.3%–96.2% of the overall BMAE reduction is already achieved within the first 20% of training. (b) The baseline models still struggle to recover local structures and fine details as training approaches its final stage. The middle-stage baseline predictions contain noticeable residual noise, blurred structures, and local artifacts, some of which persist into the late stage. Motivated by these observations, we propose reshaping training-time predictions so that, after global brightness has largely stabilized, residual local restoration errors become relatively more influential during subsequent optimization, thereby facilitating late-stage artifact suppression and structural-detail restoration. (c) Visual examples of our DITM-based models. With DITM, the same backbones produce cleaner local structures and textures with fewer residual artifacts. In this figure, red and blue boxes indicate enlarged regions. Best viewed enlarged.

Fig. 2: Training and inference pipeline of the proposed DITM. (a) Training Stage: DITM adjusts the model prediction

A

of the baseline model

f_{\theta}(\cdot)

by leveraging the global mean brightness values of the ground-truth image

Y

, generating an interpolated prediction

C

for loss computation. (b) Inference Stage: DITM is discarded, and the trained baseline

f_{\theta^{*}}(\cdot)

produces the final enhanced image

A^{*}

with no additional inference-time computation. (c) DITM: This module adaptively computes an interpolation factor

s

based on the brightness discrepancy between

A

and

Y

to yield the adjusted prediction

C

.

Datasets	Methods	Evaluation Metrics
		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DeepSSIM $\uparrow$	IQA $\uparrow$	IAA $\uparrow$
LOLv1	RetinexNet	17.076 $\pm$ 0.080	0.7095 $\pm$ 0.0020	0.424 $\pm$ 0.004	0.3834 $\pm$ 0.0028	2.032 $\pm$ 0.019	1.204 $\pm$ 0.007
	RetinexNet with DITM	17.546 $\pm$ 0.111 (+0.470)	0.7190 $\pm$ 0.0013 (+0.0095)	0.403 $\pm$ 0.015 (-0.021)	0.4002 $\pm$ 0.0208 (+0.0168)	2.144 $\pm$ 0.036 (+0.112)	1.279 $\pm$ 0.047 (+0.075)
	MIRNet	21.447 $\pm$ 0.107	0.7849 $\pm$ 0.0030	0.225 $\pm$ 0.001	0.6072 $\pm$ 0.0032	2.937 $\pm$ 0.016	1.759 $\pm$ 0.020
	MIRNet with DITM	21.703 $\pm$ 0.101 (+0.256)	0.7963 $\pm$ 0.0011 (+0.0114)	0.217 $\pm$ 0.004 (-0.008)	0.6265 $\pm$ 0.0043 (+0.0193)	2.973 $\pm$ 0.030 (+0.036)	1.750 $\pm$ 0.019 (-0.009)
	CID-Net	24.102 $\pm$ 0.184	0.8563 $\pm$ 0.0016	0.098 $\pm$ 0.003	0.8193 $\pm$ 0.0125	3.889 $\pm$ 0.048	2.097 $\pm$ 0.009

TABLE I: Quantitative comparison on LOL datasets. Each setting was independently repeated three times, and the mean and standard deviation are reported. Values in parentheses denote the mean change of DITM over the corresponding baseline. Red indicates improvement, while green indicates degradation.

Methods	DICM			LIME			MEF			NPE			VV			Weighted Average
Methods	MUSIQ $\uparrow$	IQA $\uparrow$	IAA $\uparrow$	MUSIQ $\uparrow$	IQA $\uparrow$	IAA $\uparrow$	MUSIQ $\uparrow$	IQA $\uparrow$	IAA $\uparrow$	MUSIQ $\uparrow$	IQA $\uparrow$	IAA $\uparrow$	MUSIQ $\uparrow$	IQA $\uparrow$	IAA $\uparrow$	MUSIQ $\uparrow$	IQA $\uparrow$	IAA $\uparrow$
RetinexNet	61.058	3.165	2.349	62.558	2.555	2.236	67.494	2.877	2.395	69.172	2.486	1.896	39.267	2.667	2.105	58.452	2.944	2.272
RetinexNet with DITM	61.112	3.225	2.368	61.729	2.628	2.269	67.111	2.952	2.459	68.542	2.617	2.007	37.892	2.779	2.158	58.067 ( $\downarrow$ )	3.021 ( $\uparrow$ )	2.310 ( $\uparrow$ )
MIRNet	57.611	3.551	2.473	59.912	2.816	2.292	61.546	3.251	2.609	64.569	3.156	2.329	36.278	3.127	2.084	54.748	3.350	2.395
MIRNet with DITM	58.039	3.635	2.476	59.621	2.924	2.357	61.676	3.281	2.635	64.708	3.282	2.353	37.796	3.281	2.131	55.267 ( $\uparrow$ )	3.444 ( $\uparrow$ )	2.414 ( $\uparrow$ )
CID-Net	58.324	3.619	2.399	59.359	2.939	2.397	62.508	3.463	2.598	66.436	3.475	2.436	40.264	3.606	2.230	56.081	3.534	2.396

TABLE II: Quantitative comparison on five unpaired datasets. The weighted average performance is obtained according to the image number of the five datasets.

Methods	LOLv1			LOLv2-real			LOLv2-synthetic
Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Baseline	23.498	0.8264	0.153	22.066	0.8378	0.178	25.149	0.9267	0.061
$B$ -only	20.842	0.7966	0.165	20.370	0.8005	0.334	20.462	0.8766	0.125
ISW	23.381	0.8323	0.136	22.122	0.8447	0.154	25.497	0.9278	0.059
HDI	22.413	0.8333	0.127	21.903	0.8388	0.167	25.713	0.9346	0.053
DITM (Ours)	24.050	0.8444	0.120	22.934	0.8532	0.139	25.717	0.9394	0.050

TABLE III: Ablation study of alternative interpolation strategies using RetinexFormer as the baseline.

Fig. 3: Visual comparison of different low-light image enhancement methods. From left to right: (a) Low-light Input, (b) Baseline model, (c) Ours (Baseline + DITM), and (d) Ground Truth. By integrating DITM into these existing supervised LLIE frameworks, the visual quality of the shown examples is improved in terms of artifact suppression, noise reduction, and preservation of structural details. Zoomed-in regions are provided for better visualization of local details.

Fig. 4: Visual comparison on unpaired real-world low-light datasets. From left to right: (a) Input, (b) Baseline, and (c) Ours (Baseline + DITM). The proposed DITM generally improves the generalization capability of existing supervised LLIE frameworks, resulting in improved visibility, reduced noise, and more robust restoration performance. Zoomed-in views are provided for a subset of regions for detailed comparison, while some regions are additionally highlighted using red dashed boxes to indicate noticeable artifacts.

Fig. 5: Comparison of training dynamics across different strategies on the LOLv1 training dataset. Curves of training loss, SSIM, and LPIPS are shown over 150K iterations, with values recorded every 1K iterations. Compared with the other strategies, DITM shows better performance over most of the training process.

Fig. 6: SSIM and LPIPS of RetinexFormer and CID-Net under different

\lambda

and

\delta

settings on the held-out validation subsets of LOLv1 and LOLv2-synthetic.

Fig. 7: Qualitative comparison of the enhancement performance at different training stages (10K, 80K, and 150K iterations) on the LOLv1 training dataset. Compared to the baseline, the model with DITM begins to exhibit superior detail restoration and noise suppression from the middle stage of training. Zoomed-in views are provided to better illustrate the progressive improvements across different training stages.

Fig. 8: Visualization of input gradient maps

\frac{\partial\mathcal{L}}{\partial X}

(gradients of the loss with respect to the input) and reconstruction error maps at different training stages (10K, 80K, and 150K iterations) for the Baseline (top row) and DITM (bottom row). (a) Input gradient maps suggest that DITM progressively yields more concentrated responses on structural regions, while the baseline exhibits more diffuse responses throughout training. (b) Reconstruction error maps illustrate that the errors gradually diminish during training for both methods. The model with DITM exhibits cleaner and more spatially uniform residual patterns in the later stage, suggesting a more focused training response in later stages.

Fig. 9: Distribution evolution of the adaptive interpolation factor

s

during the training of RetinexFormer on LOLv1, LOLv2-real, and LOLv2-synthetic. The distributions are recorded at 10K, 80K, and 150K training iterations. The width of each violin represents the distribution density of the per-sample interpolation factors, while the diamond marker denotes the corresponding mean value

\mu_{s}

. We additionally report

R_{s<0.1}

, which denotes the proportion of recorded samples whose interpolation factors are smaller than 0.1 at each training stage. Across all three datasets, the dominant distribution of

s

progressively shifts toward smaller values, accompanied by decreasing

\mu_{s}

and increasing

R_{s<0.1}

. The remaining thin upper tails indicate that a small number of samples retain relatively large

s

values at later stages, reflecting the per-sample adaptive behavior of DITM rather than a fixed iteration-dependent schedule.

Model	Stage	$\boldsymbol{\min(\mu_{A})}$	$\boldsymbol{\operatorname{Mean}(\mu_{A})}$	$\boldsymbol{\operatorname{Mean}(\beta)}$	$\boldsymbol{\max(\beta)}$	$\boldsymbol{\operatorname{Mean}(s)}$
RetinexFormer with DITM	Early	0.1794	0.3634	1.1964	1.7028	0.5739
	Middle	0.1896	0.4044	1.0787	1.4301	0.2617
	Late	0.1990	0.4128	1.0510	1.2256	0.1163
RetinexMamba with DITM	Early	0.1694	0.3542	1.2250	1.6734	0.6324
	Middle	0.1902	0.3923	1.1027	1.4034	0.3161
	Late	0.2064	0.4103	1.0603	1.3587	0.1354

TABLE IV: Numerical behavior of DITM on the darkest 10% samples of the LOLv1 training set at different training stages.

Methods	Evaluation Metrics
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DeepSSIM $\uparrow$	IQA $\uparrow$	IAA $\uparrow$
RetinexFormer	27.012 $\pm$ 0.607	0.9035 $\pm$ 0.0025	0.073 $\pm$ 0.002	0.8823 $\pm$ 0.0005	3.014 $\pm$ 0.009	2.176 $\pm$ 0.004
RetinexFormer with DITM	27.673 $\pm$ 0.462 (+0.661)	0.9121 $\pm$ 0.0008 (+0.0086)	0.068 $\pm$ 0.002 (-0.005)	0.8928 $\pm$ 0.0015 (+0.0105)	3.070 $\pm$ 0.040 (+0.056)	2.205 $\pm$ 0.015 (+0.029)
RetinexMamba	26.264 $\pm$ 0.577	0.9036 $\pm$ 0.0003	0.072 $\pm$ 0.001	0.8845 $\pm$ 0.0024	3.108 $\pm$ 0.016	2.256 $\pm$ 0.009
RetinexMamba with DITM	26.596 $\pm$ 0.051 (+0.332)	0.9124 $\pm$ 0.0003 (+0.0088)	0.065 $\pm$ 0.001 (-0.007)	0.8956 $\pm$ 0.0018 (+0.0111)	3.231 $\pm$ 0.022 (+0.123)	2.311 $\pm$ 0.011 (+0.055)

TABLE V: Performance comparison on the top 20% spatially non-uniform samples from the LOLv2-synthetic test set. Each setting was independently repeated three times, and the mean and standard deviation are reported. Values in parentheses denote the mean change of DITM over the corresponding baseline. Red indicates improvement, while green indicates degradation.

Methods	PSNR $\uparrow$	SSIM $\uparrow$	DeepSSIM $\uparrow$	LPIPS $\downarrow$	IQA $\uparrow$	IAA $\uparrow$
PyDiff	21.958	0.8969	0.8283	0.110	4.158	2.245
PyDiff with DITM	22.004 (+0.046)	0.8976 (+0.0007)	0.8289 (+0.0006)	0.110 (+0.000)	4.160 (+0.002)	2.241 (-0.004)
Diff-Retinex++	24.344	0.8655	0.8308	0.105	4.201	2.282
Diff-Retinex++ with DITM	24.108 (-0.236)	0.8672 (+0.0017)	0.8326 (+0.0018)	0.103 (-0.002)	4.213 (+0.012)	2.281 (-0.001)

TABLE VI: Evaluation of DITM on diffusion-based low-light image enhancement methods on the LOLv1 dataset.

Model	Training Time / Iter	Peak GPU Memory
RetinexNet	326.458 ms	2.103 GB
RetinexNet with DITM	328.990 ms (+0.78%)	2.104 GB
MIRNet	519.327 ms	20.040 GB
MIRNet with DITM	520.944 ms (+0.31%)	20.041 GB
CID-Net	210.626 ms	14.350 GB
CID-Net with DITM	213.715 ms (+1.47%)	14.375 GB

TABLE VII: Training-time overhead of DITM on the adopted LLIE backbones in our experiments.

Method	Parameter Group	Early Stage		Late Stage
Method	Parameter Group	$\rho_{g}$	$\cos$	$\rho_{g}$	$\cos$
PyDiff	Full Model	0.870	0.543	0.862	0.616
	Diffusion Denoising Module	1.000	1.000	1.000	1.000
	Image Restoration Module	0.860	0.516	0.841	0.585
Diff-Retinex++	Full Model	0.987	0.981	0.978	0.976
	Diffusion Denoising Module	1.000	1.000	1.000	1.000

TABLE VIII: Gradient response analysis of DITM on diffusion-based LLIE models. The results are averaged over sampled images.

\rho_{g}

and

\cos

denote the gradient norm ratio and cosine similarity between baseline and DITM optimization, respectively.

Figure 18

Figure 19

Figure 20

Figure 21

Figure 22

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

Abstract

Figures & tables

Explore similar work

InterLight: Leveraging Intrinsic Illumination Priors for Low-Light Image Enhancement

Internally Referenced Low-Light Enhancement

Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex