acceptodds
Under review as a conference paper at ICLR 2027

Teaching Lookup Tables to See in the Dark: Learning Complete Mappings from a Diffusion Teacher

Abstract

Low-light image enhancement (LLIE) aims to restore illumination, color, and structural edges in images captured under poor lighting. Existing LUT-based methods construct image mappings by combining basis LUTs, but are constrained by predefined transformations and supervised primarily by queries induced by the input image, leaving many grid points in high-dimensional LUTs weakly constrained. To address this limitation, this work proposes a diffusion-teacher-guided LUT distillation method for learning complete image-adaptive mappings. In the wavelet domain, the low-frequency LUT corrects brightness, contrast, and color, while the high-frequency residual LUT restores structures and suppresses noise. Specifically, target LUTs are first obtained from paired images through constrained inversion, after which a conditional diffusion teacher learns residual corrections from coarse LUTs to the target LUTs. Teacher and student interpolation outputs are then aligned using both image queries and LUT-grid queries, extending supervision to sparsely accessed regions of the LUT domain. At inference, the student predicts complete LUTs without diffusion sampling; CP decomposition provides a compact representation for lookup. Experiments on multiple benchmarks demonstrate state-of-the-art performance across a range of evaluation metrics.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-light image enhancement, Lookup table, Diffusion model, Knowledge distillation, Wavelet transform

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CWdgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a diffusion-teacher-guided method for learning complete, image-adaptive lookup tables (LUTs) aimed at low-light image enhancement (LLIE). The technique involves decomposing images using a stationary wavelet transform (SWT) and makes use of frequency-specific 5D LUTs, consisting of a LUT for the low frequencies to correct brightness and colour and another set of LUTs for the high frequencies to carry out structure restoration and noise suppression. Coarse LUTs are learned, the paired data are inverted to arrive at the target LUTs, the LUTs are then refined through the use of a conditional diffusion teacher acting on the LUT values, and the teacher's continuous interpolation function is distilled into a compact student model which is able to predict the full LUTs in a single forward pass; the CP decomposition offers a compact representation. The experiments, carried out on various benchmarks, yield strong quantitative and qualitative results together with extensive ablation studies.

Critical strengths
  • Breaks down the LLIE into frequency-specific, image-adaptive LUTs by having the high-frequency LUTs predict the residuals directly on the wavelet subbands in order to deal simultaneously with detail restoration and noise suppression.
  • It presents a diffusion teacher that works using LUT values, learning the residuals in a manner that goes from the coarse LUTs to the target LUTs, and this is a neat approach to bringing diffusion's ability to make corrections into explicit, non-iterative mappings.
  • Suggests a continuous distillation of the LUT function which monitors the student with respect to its interpolation outputs, including the use of grid-sampled in-cell queries so as to address the sparsely visited areas of the 5D domain.
Critical weaknesses
  • The target LUTs, as defined by constrained inversion (Eq. 3), may be ill-posed unless strong prior assumptions are made; the way in which Λb is constructed (specifically, how 'sparsely queried' is quantified) and the sensitivity to the hyperparameters (λa, λs, εf) are not completely specified.
  • The stability of the image-adaptive non-uniform nodes is not examined and no explicit regulariser for the smoothness of the nodes across images is mentioned.
  • It is unclear whether CP reparameterization is performed per-image at inference or precomputed/offline; per-image ALS could challenge the reported real-time claims
Questions for authors
  • What values are taken for λa, λs, and εf when the Λb weights in equation 3 are calculated in order to measure the effect of sparsely queried nodes, and can sensitivity analyses for these hyperparameters be provided?
  • With respect to continuous function distillation, how many queries per image are selected from πimg and πgrid, how is ρ determined, and how are the grid queries sampled within cells (uniformly along the axes or on the basis of importance weighting)?
  • Has any kind of regularization or stabilization been applied to the predicted non-uniform nodes? Have you come across instability as a result of rapid changes in the node arrangement?
1 comment
AuthorsOpenReview ↗

Clarifications on LUT Regularization, Query Sampling, and Node Stability

Q1: We have added "Hyperparameter Sensitivity Analysis" to Section 3.2 "Ablation Study"(Page 10), together with "Figure 6 (Page 9)", which presents the sensitivity of ten key hyperparameters.

1: Regularization weights. We set λa=0.1\lambda_a=0.1 and λs=10−2\lambda_s=10^{-2} and εf=10−5\varepsilon_f=10^{-5}. The anchoring weight λa\lambda_a controls deviations of sparsely queried nodes from the coarse LUT, while λs\lambda_s regularizes residual variations between neighboring nodes. Figures 6(8) and 6(9) show that insufficient regularization weakens mapping stability, whereas excessive regularization restricts necessary corrections. The selected values balance these effects.

2: Sensitivity analysis. We vary one parameter at a time while fixing the others. Figures 6(1)–6(2) examine LUT resolution and CP rank, showing the trade-off between mapping accuracy and parameter complexity. Figures 6(3)–6(4) show that increasing diffusion steps improves residual recovery, with diminishing gains at larger values. Figure 6(5) demonstrates the importance of balancing image and grid queries, while Figures 6(6)–6(10) examine the complementary effects of mapping regularization and function distillation.

3: Final configuration. Based on these analyses, we adopt N=33N=33, r=32r=32, K=25K=25, T=1000T=1000, ρ=0.5\rho=0.5, λsm=10−4\lambda_{\mathrm{sm}}=10^{-4}, λmon=1\lambda_{\mathrm{mon}}=1, λa=0.1\lambda_a=0.1, λs=10−2\lambda_s=10^{-2}, and λfunc=0.3\lambda_{\mathrm{func}}=0.3. This configuration achieves 27.43 dB PSNR, 0.935 SSIM, and 795K parameters, balancing mapping accuracy, regularization, and model complexity.

4: Node-wise anchoring weights. Pentalinear interpolation provides non-uniform supervision across LUT nodes. The effective supervision strength of node vv is characterized by hb(v)=∑p∈Ωwb,v(qb(p))2h_b(v)=\sum_{p\in\Omega}w_{b,v}(\mathbf{q}_b(p))^2, corresponding to a diagonal entry of the interpolation matrix Ab⊤AbA_b^\top A_b. Accordingly, Λb\boldsymbol{\Lambda}_b assigns stronger anchoring constraints to sparsely queried nodes, keeping their mappings close to the coarse LUT, while allowing sufficiently supervised nodes greater flexibility for correction. This provides complementary regularization in weakly constrained regions and improves the stability of target LUT inversion.

Q2:

1: Query allocation: For each training image and each subband, we sample 8,192 queries, comprising 4,096 image queries from πimg\pi_{\mathrm{img}} and 4,096 grid queries from πgrid\pi_{\mathrm{grid}}, yielding 32,768 queries across four subbands. Image queries are sampled uniformly without replacement from the current image's query field. We adopt fixed-ratio stratified sampling to reduce sampling variance and stabilize distillation gradients.

2: Mixing ratio ρ\rho: We have added a hyperparameter sensitivity analysis to Section 3.2 "Ablation Study" (Page 10), together with Figure 6 (Page 9). Figure 6(5) shows that both PSNR and SSIM reach their highest values at ρ=0.5\rho=0.5. Accordingly, we fix ρ=0.5\rho=0.5 across all datasets and subbands.

3: Non-uniform grid sampling: We employ stratified random sampling over image-adaptive non-uniform LUT grids without importance weighting. Specifically, we independently select an adjacent-node interval with equal probability along each of the five axes to determine a grid cell. Within the selected cell, we sample td∼U(0,1)t_d\sim\mathcal{U}(0,1) and obtain qd=νb,d(kd)+td(νb,d(kd+1)−νb,d(kd))q_d=\nu_{b,d}^{(k_d)}+t_d(\nu_{b,d}^{(k_d+1)}-\nu_{b,d}^{(k_d)}). Since the nodes are non-uniform, cell volumes vary, making the resulting sampling density inversely proportional to cell volume. Thus, each cell receives equal expected query coverage, while regions with denser nodes receive stronger supervision per unit volume. This preserves the adaptive resolution of the non-uniform LUT. Teacher and student are evaluated at identical query locations using the same 32-corner pentalinear interpolation, directly enforcing consistency between their continuous mappings.

Q3:

1: Node validity: Softmax-normalized interval proportions and cumulative sums enforce ordered nodes. An interpolation-denominator clamp of 10−610^{-6} prevents division by near-zero intervals.

2: Mapping regularization: Coarse-stage smoothness penalizes LUT slopes relative to actual node spacing, suppressing abrupt mappings across narrow intervals; monotonicity constrains low-frequency color mappings. Removing coarse smoothness reduces PSNR by 0.35 dB (Table 4).

3: Fixed-grid training: We freeze G0G_0 after coarse training. Target inversion, diffusion training, and student distillation thus share fixed nodes, preventing inter-stage grid drift.

4: Stability evidence: In a separate N=17N=17 implementation, 150 images yielded 48,000 intervals with no ordering violations. The minimum spacing was 4.9×10−34.9\times10^{-3}, 98.2% of intervals fell within 0.5--2 times uniform spacing, and input perturbations caused only 2.6×10−32.6\times10^{-3} mean node displacement. No node-induced loss spikes were observed. No explicit node-motion penalty was used.

Reviewer_wNgvOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a diffusion-teacher-guided lookup table (LUT) distillation framework for efficient low-light image enhancement. The framework first predicts coarse image-adaptive non-uniform 5D LUTs and constructs target LUTs from paired training images through constrained inversion. A conditional diffusion teacher then learns residual corrections to refine these coarse mappings. To avoid iterative diffusion sampling at inference, the refined LUT mappings are distilled into a student model using continuous function distillation with both image-derived and randomly sampled LUT-grid queries. The resulting student LUTs are further compressed through CP decomposition for efficient interpolation. Experiments across multiple paired and unpaired low-light enhancement benchmarks demonstrate strong quantitative and qualitative performance.

Critical strengths
  • Distilling from a diffusion model to a LUT-based model is an interesting idea. The authors identify that image-domain supervision touches only a limited subset of cells in a high-dimensional LUT. The paper then aligns teacher and student outputs after non-uniform pentalinear interpolation using both image queries and continuous grid queries. This is more intuitive than simply adding another loss because it directly supervises the function that will be evaluated at inference rather than only its discrete parameters.

  • Strong performance evidence. The method achieves consistently competitive results on paired and unpaired low-light enhancement benchmarks, including FiveK, LOLv1, LOLv2, and etc. The additional SMID and SDSD evaluations further demonstrate competitive restoration performance.

Critical weaknesses
  • More intuitive and direct ablation experiments design. Table 5 provides a leave-one-out style of ablation experiments that try to say that every component is backed by portion of gains. However, it says little about how and why these gains come from. The architecture ablation shows that removing teacher LUT supervision reduces PSNR by 0.39 dB (Table 5, A3), while replacing residual diffusion with complete-LUT diffusion decreases PSNR by 0.26 dB (Table 5, A7). However, neither experiment directly compares the diffusion teacher against a deterministic residual predictor with comparable capacity and conditioning. Consequently, it is unclear whether the improvements arise from diffusion modeling itself, the additional teacher capacity, or continuous-query distillation. This distinction is important because diffusion introduces substantial training complexity, and its necessity is central to the paper's claimed contribution.

  • The paper emphasizes learning complete LUT mappings, including regions sparsely covered by image-derived queries. However, the evaluation primarily measures final image reconstruction quality rather than mapping accuracy or consistency in underrepresented LUT regions. Moreover, it is unclear whether the reported runtime in Table 1 includes all necessary operations, particularly the differences among compared methods.

Questions for authors

What is the specific benefit of using a diffusion teacher instead of deterministic LUT refinement, or even other powerful architecture? Although the idea is interesting, the motivation and logical benefits should be further clarified.

4 comments
AuthorsOpenReview ↗

Response: Motivation and Benefits of Diffusion-Guided LUT Refinement

Response:

We thank the reviewer. The diffusion teacher changes the learning objective, not just capacity. Image queries leave large parts of a 5D LUT underconstrained. A deterministic regressor predicts a single correction, whereas conditional diffusion learns a prior over coherent LUT residuals. Function distillation transfers this prior to a compact student without diffusion at inference.

1: Underdetermined mappings and deterministic regression.

For subband bb, fix the non-uniform nodes NbN_b and define rb=Θbr−Θb(0)r_b=\Theta_b^r-\Theta_b^{(0)}. Since pentalinear interpolation is linear in LUT values, image-query corrections satisfy db=Abrbd_b=A_br_b. Each query accesses only 25=322^5=32 vertices. For any z∈ker⁡(Ab)z\in\ker(A_b),

Ab(rb+z)=Abrb.A_b(r_b+z)=A_br_b.

Distinct complete LUTs can therefore produce identical reconstructed images. Constrained inversion (Eq. 3) stabilizes the solution through coarse-LUT anchoring and smoothness, but these penalties do not learn joint residual statistics.

Let Cb=(Iin,Nb,Θb(0),b)C_b=(I_{\mathrm{in}},N_b,\Theta_b^{(0)},b) and RbR_b denote the condition and target residual. Under MSE, the population-optimal direct regressor predicts

f∗(Cb)=E[Rb∣Cb].f^*(C_b)=\mathbb E[R_b\mid C_b].

For two equally plausible corrections, the prediction (R(1)+R(2))/2(R^{(1)}+R^{(2)})/2 can blend incompatible cross-node patterns. Increasing CNN or Transformer capacity improves approximation of this predictor but does not change its MSE-optimal target or remove the image-query null space. The key distinction is therefore between estimating one residual and learning its conditional joint distribution.

2: Conditional diffusion as a learned residual prior.

We diffuse LUT residuals, not enhanced images. With Rb=Θbr−Θb(0)R_b=\Theta_b^r-\Theta_b^{(0)}, the forward process is

Zτ=αˉτRb+στϵ,στ=1−αˉτ,ϵ∼N(0,I).Z_\tau=\sqrt{\bar\alpha_\tau}R_b+\sigma_\tau\epsilon,\quad \sigma_\tau=\sqrt{1-\bar\alpha_\tau},\quad\epsilon\sim\mathcal N(0,I).

The teacher minimizes

Ldiff=Eτ,ϵ,Rb∥Dϕ(Zτ,τ,Cb)−ϵ∥F2.\mathcal L_{\mathrm{diff}}=\mathbb E_{\tau,\epsilon,R_b}\|D_\phi(Z_\tau,\tau,C_b)-\epsilon\|_F^2.

At the population optimum,

D∗(z,τ,Cb)=E[ϵ∣z,Cb]=−στ∇zlog⁡pτ(z∣Cb).D^*(z,\tau,C_b)=\mathbb E[\epsilon\mid z,C_b]=-\sigma_\tau\nabla_z\log p_\tau(z\mid C_b).

Thus, multi-level denoising estimates the conditional score field rather than only its mean. Reverse sampling follows this learned field toward residual configurations supported by the training distribution. Because RbR_b contains the entire LUT, it models dependencies across nearby and distant nodes beyond local smoothness. Image, node, coarse-LUT, and subband conditioning adapts corrections to the scene and frequency. The coarse LUT supplies the dominant transformation; diffusion refines the remaining residual without moving the nodes. LL handles illumination, color, and contrast; LH/HL/HH address texture and noise.

A larger deterministic predictor could also learn cross-node correlations, but under direct MSE supervision it still optimizes a point estimate. Our teacher instead learns a distribution-level correction prior during training, then transfers its functional effect to a lightweight student. This separates the capacity used to learn structured corrections from the cost of deployment.

3: Continuous function distillation.

Conditional DDIM generates Θbt=Θb(0)+R^b\Theta_b^t=\Theta_b^{(0)}+\widehat R_b. With fixed teacher targets, the student minimizes

Lfunc=Eq∼πb∥Fb(q;Θbs)−Fb(q;Θbt)∥22,\mathcal L_{\mathrm{func}}=\mathbb E_{q\sim\pi_b}\|F_b(q;\Theta_b^s)-F_b(q;\Theta_b^t)\|_2^2,

where πb=ρπbimg+(1−ρ)πbgrid\pi_b=\rho\pi_b^{\mathrm{img}}+(1-\rho)\pi_b^{\mathrm{grid}}. Image queries preserve observed corrections; grid queries extend teacher supervision to unvisited regions. For fixed nodes, Fb(q;Θb)=w(q)⊤ΘbF_b(q;\Theta_b)=w(q)^\top\Theta_b. Writing the corresponding interpolation matrices as AimgA_{\mathrm{img}} and AgridA_{\mathrm{grid}} gives

H=ρAimg⊤Aimg+(1−ρ)Agrid⊤Agrid,H=\rho A_{\mathrm{img}}^\top A_{\mathrm{img}}+(1-\rho)A_{\mathrm{grid}}^\top A_{\mathrm{grid}},

ker⁡(H)=ker⁡(Aimg)∩ker⁡(Agrid).\ker(H)=\ker(A_{\mathrm{img}})\cap\ker(A_{\mathrm{grid}}).

Grid queries shrink the null space when they constrain previously unseen directions. Diffusion thus supplies image-conditioned cross-node targets; function distillation transfers them as continuous mapping constraints rather than merely matching LUT vertices. This constrains mappings beyond the current image queries.

4: Evidence and efficiency.

Table 5 directly tests residual diffusion: removing it (A7) lowers PSNR from 27.43 to 27.17 dB and SSIM from 0.935 to 0.931, while raising ΔE\Delta E from 6.75 to 6.96. Direct supervision with inverted target LUTs (A3) achieves 27.04 dB/0.928/7.05, versus 27.43 dB/0.935/6.75 for the full model, supporting the combined refinement-and-distillation design. Removing grid queries (A6) costs 0.29 dB PSNR; removing function distillation (Table 4) costs 0.79 dB. A7 isolates diffusion; A3 tests the broader pipeline.

Diffusion is restricted to teacher training and target generation. The deployed student predicts LUTs in one pass, with 795K parameters and 3.5 ms inference, without denoising.

AuthorsOpenReview ↗

Response: Motivation and Benefits of Diffusion-Guided LUT Refinement

We thank the reviewer for this question. The fundamental distinction lies in the learning objectives: direct regression predicts a single LUT correction, whereas conditional diffusion learns a distribution over residual corrections to a coarse LUT. This distinction is particularly important under incomplete image supervision.

1. Why is direct LUT regression challenging?

With fixed non-uniform nodes, each five-dimensional query interpolates 32 vertices in one cell. Image reconstruction thus constrains the LUT at pixel-induced queries, leaving many cells unvisited. Different LUTs may produce identical enhanced images while giving different outputs elsewhere. Constrained inversion (Eq. 3) addresses this ambiguity by anchoring sparsely queried nodes to the coarse LUT and regularizing residual variations, yielding stable target mappings for subsequent prior learning.

Under squared-error supervision, the optimal deterministic prediction is the conditional mean of the target residual. When observations admit multiple plausible residual patterns, averaging them may weaken coordinated variations across LUT nodes. A larger CNN or Transformer can represent more complex dependencies, but greater capacity alone does not change this point-estimation objective. We therefore seek to learn a conditional distribution of residual corrections, not merely increase model capacity.

2. What does the conditional diffusion teacher contribute?

We first construct target LUTs by constrained inversion and define the residual between each target LUT and its coarse prediction (Eqs. 3--5). Conditioned on the input image, fixed non-uniform nodes, coarse LUT, and subband identity, the teacher denoises the complete residual table at multiple noise levels. This training models plausible joint variations among LUT nodes, including dependencies not captured by local smoothness alone. DDIM then generates a residual correction, which is added to the coarse LUT to obtain the teacher mapping.

The coarse LUT supplies the dominant transformation, allowing diffusion to focus on remaining errors. Fixed nodes keep coarse, target, and teacher mappings in the same query coordinates. Subband conditioning distinguishes low-frequency illumination, color, and contrast correction from high-frequency texture recovery and noise suppression. The teacher thus learns a scene-dependent prior over LUT residuals, capturing coordinated correction patterns that guide the mapping across sparsely observed regions. Through distribution-oriented, multi-noise-level training, the diffusion teacher models joint residual dependencies across LUT nodes, providing a richer learning objective than direct point regression.

3. Why is continuous function distillation needed?

The teacher's corrections must reach the student without diffusion at inference. Matching enhanced images alone would restrict supervision to the queries present in the input. We instead match teacher and student outputs after non-uniform pentalinear interpolation at both image-derived and continuous grid queries (Eqs. 6--7). Image queries retain fidelity on the observed distribution; grid queries convey teacher-defined corrections to regions the current image may not visit. This supervision targets the function actually executed by the LUT, rather than only its discrete node values.

The two mechanisms therefore have distinct roles: diffusion learns structured residual corrections, while function distillation transfers their effect over a broader query domain. Grid queries extend the teacher's learned correction knowledge beyond the observed image distribution. The student subsequently predicts its LUTs in a single forward pass.

4. What do the experiments show?

Table 5 distinguishes residual diffusion from full-LUT diffusion: replacing the former with the latter (A7 vs. A10) reduces PSNR from 27.43 to 27.17 dB and SSIM from 0.935 to 0.931, while increasing ΔE\Delta E from 6.75 to 6.96. Replacing teacher refinement with direct target-LUT supervision (A3 vs. A10) decreases PSNR by 0.39 dB; removing grid queries (A6 vs. A10) decreases it by 0.29 dB. In Table 4, removing function distillation (S3 vs. S8) costs 0.79 dB. These results consistently demonstrate the contributions of structured residual refinement, teacher-guided supervision, and expanded query coverage, supporting the proposed framework as a coherent approach to learning image-adaptive LUT mappings.

Finally, diffusion is confined to training and teacher-LUT generation. The deployed student has 795K parameters and a reported 3.5 ms inference time, with no diffusion sampling. We thus use diffusion to learn structured corrections under incomplete supervision, then transfer them to an efficient deterministic LUT predictor. We hope this clarifies the rationale and methodological value of the diffusion teacher.

Thanks for your response! I appreciate the author’s time and effort, though I still hope the AC will carefully consider the feedback from other active reviewers (if any exists).

I replied in the wrong place just now—sorry about that. The correct location should be below my reply.

Thanks to the authors for their reply and sorry for the delayed response. But, it seems to me that the authors didn't respond to my concerns on more intuitive and direct ablation experiments and explanations on the report of latency. Instead, they emphasize the motivation and benefits. Thus, unfortunately, the reply can't convince me, and I will maintain my original rating.

Reviewer_ZwUBOpenReview ↗
Rating
2
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The goal of low-light image enhancement is to brighten dark images, restore accurate colors, and reveal fine details—all while ensuring fast inference. Previous methods based on look-up tables (LUTs) were limited to brightness and color adjustments and could not address image noise; furthermore, since these LUTs were composed of multiple fixed templates, many regions remained unrepresented by the training data. Our approach consists of three steps. First, the image is decomposed into low-frequency and high-frequency components: a low-frequency LUT handles brightness and color, while a high-frequency LUT manages detail restoration and noise suppression. Second, a diffusion model is trained as a "teacher" to learn how to fill in the gaps in the LUTs where training data is absent. Third, this knowledge is distilled from the teacher into a "student" network. During inference, the student network generates the complete LUTs in a single forward pass—eliminating the need to run the diffusion model—and processes each image in just 3.5 milliseconds. On standard datasets such as FiveK and LOL, our method achieves state-of-the-art performance, outperforming the next-best approach by 1.26 to 1.90 dB in terms of PSNR and other metrics.

Critical strengths

The pipeline design is comprehensive and self-consistent, incorporating frequency-split LUTs, non-uniform nodes, constraint inversion, diffusion-based teacher models, and continuous function distillation—with ablation studies provided for each component. The results are robust and consistent: the method significantly outperforms the runner-up on the LOL dataset (across all three subsets) and ranks first in 13 no-reference benchmark metrics, all while maintaining competitive efficiency.

The continuous function distillation mechanism is highly compelling; by performing continuous sampling within the grid to supplement sparse supervision, it offers a sound solution to the challenges of sparse supervision in high-dimensional LUTs. The 0.40 dB improvement of S8 over S5 demonstrates that these gains stem from the mechanism itself rather than merely an increase in model capacity.

Critical weaknesses

The core claims lack sufficient evidence. There is no direct quantitative analysis of the sparse-query regions; the contribution of the "diffusion teacher" is not isolated from that of a standard regressor under equivalent conditions; and the gains attributed to the teacher model are confounded by increased model capacity and extended training stages. Furthermore, sensitivity analyses regarding the number of DDIM steps and the CP rank are missing.

The evaluation design limits the validity of the SOTA (state-of-the-art) claims. The approach relies on paired data; for unpaired scenarios, it provides only reference-free metrics and lacks user studies. Additionally, while the main text mentions approximately 20 baselines, only 7 are reported for the FiveK dataset.

Questions for authors

Could you provide direct evidence of the complete mapping—specifically, by analyzing the error across the LUT query domain partitions, to clarify which grid cells are missed during the image lookup process, and how the distillation mechanism compensates for these gaps?

Regarding the marginal gains of the diffusion-based approach over a direct regressor (under equivalent experimental conditions), does it offer specific advantages for sparse regions, and how do these gains evolve as the number of DDIM steps varies?

Why are only seven baselines reported for the FiveK dataset? Could you include results for Adair, CIDNet, DnLUT, and FLUT? Furthermore, were official checkpoints used consistently across all these models?

4 comments
AuthorsOpenReview ↗

Response to Reviewer: LUT Coverage, Diffusion Distillation, and Comparative Evaluation

Q1:

We thank the reviewer. We analyze query gaps and compensation.

1: Coverage and identifiability: Non-uniform nodes partition each 5D LUT; counts reveal unvisited cells. With fixed nodes, pentalinear interpolation defines AimgA_{\mathrm{img}}, with 32 vertices per query. A vertex vv outside all image-query supports satisfies Aimgev=0A_{\mathrm{img}}e_v=0. Changing it leaves reconstruction unchanged; thus image supervision cannot identify the full LUT.

2: Distillation compensation: Grid queries introduce additional constraints:

H=ρAimg⊤Aimg+(1−ρ)Agrid⊤Agrid,H=\rho A_{\mathrm{img}}^\top A_{\mathrm{img}}+(1-\rho)A_{\mathrm{grid}}^\top A_{\mathrm{grid}},

ker⁡(H)=ker⁡(Aimg)∩ker⁡(Agrid).\ker(H)=\ker(A_{\mathrm{img}})\cap\ker(A_{\mathrm{grid}}).

This reduces the null space in newly constrained directions. With independent uniform sampling inside a normalized 5D cell, the interpolation-weight second moment has minimum eigenvalue 6−56^{-5}, implying

Ec≥6−5∣Θcs−Θct∣F2.\mathcal E_c\geq6^{-5}\\|\Theta_c^s-\Theta_c^t\\|_F^2.

This constrains all 32 vertex differences in expectation. The bound is not a finite-sample guarantee. For a grid query qq in an image-unvisited cell,

ℓc(q)=∣Fs(q)−Ft(q)∣22,\ell_c(q)=\\|F_s(q)-F_t(q)\\|_2^2,

∇Θc,asℓc(q)=2wa(q)[Fs(q)−Ft(q)].\nabla_{\Theta_{c,a}^s}\ell_c(q)=2w_a(q)[F_s(q)-F_t(q)].

Positive interior weights propagate nonzero discrepancies to all vertices. Coarse-LUT anchoring and smoothness regularize sparse regions; the diffusion teacher learns residual corrections conditioned on the image, nodes, and coarse LUT. Grid queries transfer these targets without requiring ground-truth LUT values at unvisited locations.

3: Evidence: Grid queries improve PSNR/SSIM by 0.29 dB/0.005 and reduce ΔE\Delta E by 0.23 (Table 5); removing function distillation reduces PSNR by 0.79 dB (Table 4). Errors in visited, adjacent-unvisited, and other unvisited cells can quantify compensation.

Q2:

We address diffusion gains, sparse regions, and DDIM steps.

1: Direct regression vs. diffusion: Constrained inversion yields target LUTs; diffusion learns residuals relative to coarse LUTs. Let Cb=(Iin,Nb,Θb(0),b)C_b=(I_{\mathrm{in}},N_b,\Theta_b^{(0)},b) and Rb=Θbr−Θb(0)R_b=\Theta_b^r-\Theta_b^{(0)}. Under MSE, the optimal direct regressor predicts

fω∗(Cb)=E[Rb∣Cb].f_\omega^*(C_b)=\mathbb E[R_b\mid C_b].

The diffusion teacher instead learns conditional residual distributions through multi-noise-level denoising:

Ldiff=Eτ,ϵ,Rb[∣Dϕ(Zτ,τ,Cb)−ϵ∣F2],\mathcal L_{\mathrm{diff}}=\mathbb E_{\tau,\epsilon,R_b}\left[\\|D_\phi(Z_\tau,\tau,C_b)-\epsilon\\|_F^2\right],

where Zτ=αˉτRb+1−αˉτϵZ_\tau=\sqrt{\bar\alpha_\tau}R_b+\sqrt{1-\bar\alpha_\tau}\epsilon. This captures cross-node dependencies for difficult corrections; the coarse LUT already handles the main transformation.

2: Sparse-region advantages: Image queries unevenly cover the 5D LUT. The effective image supervision of vertex vv is

hv=∑q∈Qimgwv(q)2.h_v=\sum_{q\in Q_{\mathrm{img}}}w_v(q)^2.

Small hvh_v means weak supervision; hv=0h_v=0 means no image-query support. Anchoring and smoothness stabilize these mappings, while diffusion learns correlated correction residuals. Grid queries transfer teacher targets through function distillation:

Lfunc=Eq∼πb∣Fb(q;Θbs)−Fb(q;Θbt)∣22.\mathcal L_{\mathrm{func}}=\mathbb E_{q\sim\pi_b}\\|F_b(q;\Theta_b^s)-F_b(q;\Theta_b^t)\\|_2^2.

Teacher targets remain fixed during student training. Thus, these priors constrain unobserved locations.

3: DDIM steps: With KK steps, Θb,Kt=Θb(0)+R^b,K\Theta_{b,K}^t=\Theta_b^{(0)}+\widehat R_{b,K}. Define the gain over direct regression as Δ(K)=PSNRdiff(K)−PSNRreg\Delta(K)=\mathrm{PSNR}_{\mathrm{diff}}(K)-\mathrm{PSNR}_{\mathrm{reg}}. Expected function errors obey

Estudent(K)≤Edist(K)+Eteach(K),\sqrt{E_{\mathrm{student}}(K)}\leq\sqrt{E_{\mathrm{dist}}(K)}+\sqrt{E_{\mathrm{teach}}(K)},

where the terms on the right measure student–teacher and teacher–target differences under the same query distribution. More steps can reduce discretization error, but denoiser and distillation errors limit further gains; improvements may saturate. Teacher DDIM steps do not affect student inference. Relative to A3 (direct target-LUT supervision), our full method gains 0.39 dB PSNR and 0.007 SSIM and reduces ΔE\Delta E by 0.30 (Table 5).

Q3:

1: Expanded comparisons: Table 1 adds Adair, CIDNet, DnLUT, and FLUT, reporting PSNR, SSIM, ΔE\Delta E, LPIPS, parameters, and runtime at 4K/480p. At 4K, ours achieves 27.25 dB PSNR and 0.946 SSIM, exceeding FLUT by 0.68 dB and SvdLUT by 0.008. Its ΔE\Delta E/LPIPS (6.62/0.07) improve by 0.52/0.02. At 480p, ours reaches 27.44 dB, 0.952 SSIM, 6.73 ΔE\Delta E, and 0.05 LPIPS, improving by 0.56 dB/0.006/0.25/0.02 over the best competing values. It uses 795K parameters and 3.5 ms versus FLUT's 2,080K and 5.3 ms. Runtime excludes teacher DDIM sampling.

2: Checkpoint consistency: We retrained the competing models locally using consistent dataset splits, training settings, and evaluation procedures while preserving their architectures. All results use retrained checkpoints on identical test sets, with consistent preprocessing and metric computation.

AuthorsOpenReview ↗

Response to Reviewer: LUT Coverage, Diffusion Distillation, and Comparative Evaluation

We thank the reviewer for these questions. We distinguish what the current experiments establish from the behavior that remains to be measured.

1.LUT coverage and compensation for missing queries.

A five-dimensional LUT query interpolates only 32 vertices in its containing cell. Image reconstruction therefore constrains the mapping only where the input generates queries; an unvisited cell receives no direct pixel-query supervision, although it may share vertices with visited cells. In a separate coverage diagnostic on 15 full-resolution SICE images using 17 uniformly spaced nodes per axis, the proportions of cells receiving no pixel queries were 99.556% (LL), 99.953% (LH), 99.944% (HL), and 99.991% (HH). These numbers describe the diagnostic grid, not the final 33-node LUT. In one LL image, adjacent cells (0,1,3,1,0) and (0,0,3,1,0) received 468,693 and zero queries. They show why image reconstruction cannot validate a complete mapping; "complete" means defined everywhere, not ground-truth verified.

The mechanism is straightforward. For fixed nodes, interpolation is linear in LUT values; any vertex not involved in an image query can change without affecting the reconstructed image. Constrained inversion (Eq. 3) anchors weakly supported values to the coarse LUT and regularizes neighboring residuals, but it cannot supply missing ground truth. Our grid queries (Eqs. 6–7) instead evaluate the teacher and student interpolants inside LUT cells, extending teacher supervision beyond the input's pixel queries. This gives a learning signal for mapping locations absent from the image, without pretending that teacher outputs are ground-truth LUT values. Removing grid queries reduces PSNR from 27.43 to 27.14 dB (Table 5, A10 vs. A6); removing function distillation reduces it to 26.64 dB (Table 4, S8 vs. S3). These ablations show a benefit, but do not quantify error separately in visited and unvisited cells. Such a direct analysis requires held-out continuous queries grouped by cell occupancy and comparison against the same teacher for students trained with and without grid queries.

2.Diffusion versus direct regression; sparse regions and DDIM steps.

A deterministic model trained by residual MSE predicts a point estimate conditioned on the image, nodes, and coarse LUT. If several residual patterns are plausible, its squared-error optimum is their conditional mean, potentially averaging incompatible cross-node corrections. Conditional diffusion instead learns the residual distribution through denoising at multiple noise levels (Eqs. 4–5), which can represent joint corrections across LUT nodes. The coarse LUT carries the main transformation; diffusion models what remains. This is a principled motivation, not a claim that sufficiently powerful deterministic networks cannot learn such dependencies.

The potential advantage is most relevant where image-query constraints are weak. The diffusion teacher supplies image-conditioned residual targets, and continuous grid queries transfer them to the student in those regions. Existing ablations support the overall design: direct supervision with inverted target LUTs (A3) gives 27.04 versus 27.43 dB (Table 5, A3 vs. A10), while diffusing full LUTs rather than residuals gives 27.17 dB (A7 vs. A10). A3 is not a capacity-matched deterministic alternative, so neither comparison alone proves diffusion's superiority specifically in sparse cells. Figure 6(3) shows that increasing DDIM steps improves PSNR/SSIM, with diminishing gains around the chosen K=25; these steps are used to generate teacher LUTs and do not affect student inference. A step-by-step advantage over a fixed direct regressor, and sparse-cell error curves, would require additional matched measurements beyond the current figure.

3.FiveK comparisons and checkpoint protocol.

We have expanded Table 1 to include Adair, CIDNet, DnLUT, and FLUT, with PSNR, SSIM, ΔE, LPIPS, parameters, and runtime at both 4K and 480p. At 4K, our method achieves 27.25 dB PSNR, 0.946 SSIM, 6.62 ΔE, and 0.07 LPIPS: 0.68 dB above FLUT in PSNR, 0.008 above SvdLUT in SSIM, and 0.52/0.02 below FLUT in ΔE/LPIPS. At 480p, it achieves 27.44 dB, 0.952, 6.73, and 0.05, improving on the strongest competing values by 0.56 dB, 0.006, 0.25, and 0.02, respectively. The deployed student has 795K parameters and a reported runtime of 3.5 ms; no teacher DDIM steps are performed at inference.

Regarding checkpoints, we did not mix results from official pretrained weights with locally reproduced models. The reported comparison uses locally retrained checkpoints, with common dataset splits, training protocols, preprocessing, and metrics, while retaining each baseline's architecture. We will make this distinction explicit so that the results are understood as a controlled retraining comparison rather than evaluations of official checkpoints.

The author has addressed some of my concerns; however, as I am not very familiar with this field, the technical novelty of the paper appears to be borderline in my view. Furthermore, the writing and the presentation of the figures are somewhat disorganized. If I had to assign a score, it would be around 2.4–2.5—not quite a "weak reject." I hope other reviewers will participate actively in the discussion, as I am always open to learning something new. Feel free to assign a lower weight to my assessment.

AuthorsOpenReview ↗

Further Clarification of Technical Novelty

We thank the reviewer for the follow-up and for considering our responses. We further clarify the underlying learning problem and our methodological contribution.

1. From Image Reconstruction to Mapping Identifiability

LUT methods typically learn transformations through image reconstruction. Yet reconstructing an image does not determine its underlying mapping. Each pixel constrains only the vertices involved in its interpolation; vertices outside all image-query supports remain unconstrained. With fixed nodes, Aimgev=0A_{\mathrm{img}}e_v=0 for such a vertex vv. Distinct LUTs can therefore produce the same reconstructed image while behaving differently at unseen queries. Low reconstruction error does not imply a well-determined function.

This is acute for high-dimensional adaptive LUTs. Each image requires its own mapping, but its pixels cover only part of the query domain. Increasing capacity does not create missing observations; non-uniform nodes redistribute resolution, not supervision. Our central problem is how to learn coherent, continuous correction functions from incomplete image supervision.

2. From Reconstruction Constraints to Conditional Mapping Priors

We construct target LUTs through constrained inversion, jointly enforcing reconstruction fidelity, coarse-LUT anchoring, and smoothness. This selects stable mappings from an underdetermined space for teacher learning.

Conditional diffusion then learns residual corrections relative to the coarse LUT, conditioned on the image, node positions, and initial mapping. The coarse LUT captures the main transformation; the teacher models dependencies among residual corrections. Multi-noise-level learning captures cross-node correction patterns as an image-conditioned prior.

Unlike image-space diffusion, the teacher models correction functions rather than generating restored pixels. Its prior links local observations to broader mapping structure and guides corrections in sparsely sampled regions. Inversion constructs regularized targets; diffusion learns transferable residual structure.

3. From Discrete Node Matching to Continuous Function Supervision

We next transfer this mapping behavior to a compact student. Image reconstruction supervises observed queries; node matching penalizes discrepancies between LUT entries. We align teacher and student interpolation outputs at image queries and additional continuous queries inside LUT cells. Teacher-guided supervision thereby extends beyond locations visited by the input image.

This changes both what is transferred and where it is supervised. The student learns a correction function, not merely node values or enhanced pixels. Under limited approximation capacity, node-value error and interpolation-output error induce different objectives. Our loss, Eq∼π∥Fs(q)−Ft(q)∥22\mathbb{E}_{q\sim\pi}\|F_s(q)-F_t(q)\|_2^2, directly optimizes the mapping executed at inference. Additional grid queries also convey the teacher's learned prior to regions absent from the current image.

Continuous function distillation therefore connects sparse image observations with supervision over a broader mapping domain, while preserving efficient LUT inference without diffusion sampling.

4. Frequency-Specific Mappings Follow Degradation Structure

Low-light photometric and structural degradations require different correction cues. Similar RGB values may need different adjustments under different illumination or contrast, whereas texture and noise depend on neighborhood structure.

We therefore define frequency-specific LUTs in the wavelet domain. The low-frequency LUT conditions on color, illumination, and contrast; the high-frequency LUTs use neighboring structural responses to predict residual corrections while retaining the original coefficients. These representations support inversion, prior learning, and distillation.

5. A Unified Learning Formulation

The framework follows a coherent progression: frequency-specific queries define correction spaces; constrained inversion constructs regularized targets; conditional diffusion learns residual mapping priors; continuous-query distillation transfers correction behavior beyond observed pixels; and the student executes explicit LUTs efficiently.

The advance is not a combination of familiar modules, but a shift from image-only reconstruction toward prior-guided learning of continuous, image-adaptive LUT functions. This bridges sparse observations and a broader correction domain, with implications for efficient high-dimensional mapping.

We will clarify this logic in revised figures and a reorganized method section, distinguishing training from inference.

We appreciate the reviewer's constructive engagement and hope these clarifications support reconsideration of the paper's novelty and overall contribution.

Comments on the paper

1 comment

We thank the reviewer for the reassessment. We further clarify that our contribution lies in the learning target, mapping representation, and supervision: learning an image-adaptive correction function from limited observations.

1. The underlying challenge is partial observability of LUT mappings.

Conventional LUT learning relies on image reconstruction, which constrains mappings only at pixel-induced queries. Each five-dimensional interpolation accesses 32 vertices in one grid cell, whereas an image covers only a small fraction of the domain. Thus, low reconstruction error does not ensure reliable behavior elsewhere: different LUTs can yield identical enhanced images but different outputs at unvisited locations. Because each input generates its own LUT, other images cannot directly determine its unvisited regions. Non-uniform nodes improve allocation but cannot eliminate this supervision gap.

2. The diffusion teacher learns corrections to an explicit mapping, not restored pixels.

Constrained inversion (Eq. 3) constructs target LUTs from paired images, anchoring sparsely queried nodes to the coarse LUT and regularizing neighboring residuals. Conditional diffusion then models coarse-to-target LUT residuals (Eqs. 4–5), conditioned on the input image, non-uniform nodes, coarse LUT, and subband identity. Unlike direct image restoration or independent LUT generation, it refines an existing correction function while retaining its coarse mapping. Diffusion updates LUT values, not node positions, ensuring common query coordinates across coarse, target, teacher, and student LUTs.

3. Continuous function distillation changes what and where we supervise.

Image-output distillation evaluates observed pixels, while node matching compares discrete LUT values. We instead align teacher and student correction functions defined by non-uniform pentalinear interpolation. Eq. 6 combines image and continuous grid queries; Eq. 7 matches their interpolation outputs. Image queries preserve fidelity to the observed distribution; grid queries extend teacher supervision to regions the current image may not visit. Different queries activate different interpolation vertices and weights, constraining mapping behavior beyond image reconstruction.

Continuous function distillation directly constrains interpolation behavior under the query distribution, aligning supervision with the mapping executed at inference. Under finite capacity and imperfect optimization, minimizing node errors and interpolation errors over a query distribution emphasizes different mapping properties. The novelty lies in function-level knowledge transfer over an expanded domain, where the student learns the teacher's continuous correction behavior.

4. The mapping representation reflects low-light degradation.

Rather than distilling a generic color LUT, we learn frequency-specific corrections. The low-frequency five-dimensional query combines RGB coefficients, local illumination, and contrast, so similar colors under different lighting receive different adjustments. High-frequency LUTs use signed structural responses at a center and four neighbors, predicting residuals that preserve original wavelet coefficients while restoring detail and suppressing noise. Together, the branches address complementary photometric and structural degradation.

5. Ablations support these distinctions.

Replacing function distillation with node-value matching lowers PSNR by 0.40 dB (Table 4, S5 vs. S8). Removing grid queries lowers it by 0.29 dB (Table 5, A6 vs. A10); replacing teacher refinement with direct target-LUT supervision lowers it by 0.39 dB (A3 vs. A10); and diffusing full LUTs instead of residuals lowers it by 0.26 dB (A7 vs. A10). These results demonstrate the complementary contributions of function-level supervision, expanded query coverage, and residual refinement to learning coherent correction functions across the LUT domain.

Finally, "complete mapping" refers to a student LUT that defines a continuous interpolated correction function throughout the query domain, with its behavior jointly shaped by inversion regularization, learned teacher corrections, and continuous grid-query supervision. These mechanisms extend the learned correction relationships from observed image queries to the broader mapping domain.

In summary, we move beyond fitting observed image outputs toward learning explicit image-adaptive correction functions under partial observation. Diffusion refines the LUT mapping; continuous function distillation transfers that correction over a broader domain, enabling inference without iterative diffusion. The contribution is a unified learning framework that extends LUT supervision from observed image outputs to continuous, image-adaptive correction functions. We hope this further clarifies the methodological novelty and supports reconsideration of the paper's overall contribution.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.