acceptodds
Withdrawn from ICLR 2027

GICE: Generalized Image Contrast Enhancement with Diffusion Priors

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world photographs often exhibit mixed contrast degradations induced by challenging illumination conditions (e.g., low-light, backlit, and over-exposure) together with local corruptions such as noise and compression, which reduce visibility and obscure fine details. Most existing methods target a single specific enhancement setting, often leading to brittle behavior when multiple degradations co-occur within the image. In this work, we present Generalized Image Contrast Enhancement (GICE), a unified framework that addresses diverse contrast degradations while recovering realistic textures and details. Specifically, we cast generalized contrast enhancement as a conditional image generation problem: we condition a pre-trained diffusion model on a degradation-robust representation to enable stable enhancement under heterogeneous contrast degradations. This representation is learned through multi-exposure supervision and structure-related corruption augmentation, which disentangles scene content from illumination variations and local distortions, thereby providing a stable, task-agnostic condition. Given this condition, we train the GICE model by finetuning on the pre-trained Stable Diffusion model in an end-to-end manner for efficient one-step refinement, improving image local contrast as well as fine details. Extensive experiments demonstrate that GICE generalizes well across diverse lighting conditions and consistently outperforms previous task-specific models. The code and models of GICE will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image enhancement

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QqsUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GICE, a unified diffusion-based framework for generalized image contrast enhancement, aiming to address diverse illumination degradations, including low-light, backlit, and overexposed conditions. Extensive experiments across low-light enhancement, exposure correction, and backlit enhancement demonstrate competitive performance against task-specific and unified image enhancement methods, with favorable inference efficiency.

Critical strengths

(1). The paper addresses an important limitation of existing image enhancement methods, which are often optimized for individual degradation types and may struggle when multiple degradations coexist.

(2). The proposed framework combines exposure-aware decomposition with layer-adaptive diffusion conditioning. Separating exposure-related appearance information from relatively stable structural content provides an intuitive motivation for handling contrast correction and detail reconstruction jointly.

Critical weaknesses

(1). The central assumption of the method is that the input image can be decomposed into an exposure-variant component EE and an exposure-invariant residual C=I−EC=I-E. However, this decomposition is fundamentally underconstrained. The consistency loss only encourages the residuals from different exposures to be similar, without guaranteeing that they represent genuinely exposure-invariant structures. Moreover, severe under-exposure and saturation can significantly alter or eliminate observable textures, making strict pixel-wise consistency across exposure levels questionable. The decomposition visualizations in Fig. 12 primarily suggest a separation between low-frequency appearance and high-frequency structures, which could potentially be achieved using conventional frequency decomposition or low-pass filtering..

(2). The overall framework combines several established techniques, including multi-exposure supervision, additive image decomposition, pretrained diffusion models, layer-wise feature injection, LoRA adaptation, and VAE skip connections. Although their integration is reasonable, the technical distinction from existing diffusion-based enhancement frameworks remains insufficiently demonstrated. In particular, the layer-adaptive conditioning mechanism largely follows a conventional attention-based fusion and residual injection formulation.

(3). The training pipeline selects pseudo-ground-truth images using no-reference image quality assessment metrics, including ARNIQA, MUSIQ, CLIP-IQA+, and MANIQA. However, these same metrics are subsequently used as major criteria for evaluating the proposed method. This introduces a potential evaluation bias, as the training targets are explicitly selected according to the perceptual preferences of the evaluation metrics.

Questions for authors

The paper addresses a practically important problem and presents a reasonably designed unified framework for image contrast enhancement. The extensive quantitative evaluations, qualitative comparisons, and favorable inference efficiency demonstrate promising empirical performance across different enhancement tasks. Nevertheless, I have several concerns regarding the methodological novelty and experimental validity.

Reviewer_B1JcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles a unified contrast enhancement problem, handling low-light, backlit, and exposure correction all together, even when images also have noise, blur, or JPEG artifacts. The main idea is straightforward:

The method has a decomposition module that splits the input into two parts: one that changes with exposure (E) and one that stays stable (C). Then it feeds both parts into a fine-tuned SD-Turbo model, where different layers can pick which part they need. Training uses about 48K multi-exposure image groups from public RAW datasets, with pseudo ground-truth made by fusing multiple exposures and picking the best one using IQA scores.

They test on several datasets (LSRW, UHDLL, SICE, BAID, Backlit) and compare against both task-specific methods and recent unified models. To make things fair, they retrain baselines on the same data and scale them to similar size (~1.2B parameters). Some baselines also get paired with PiSA-SR for detail recovery.

Critical strengths
  1. The problem is real and well-motivated: Real photos do have these mixed issues—dark regions with noise, bright blown-out areas, JPEG artifacts all at once. Fig. 1 shows this clearly, and it's obvious that single-task methods struggle here. The motivation for a unified approach makes sense.
  2. Fair comparison effort: I appreciate that you retrained baselines on the same data and scaled them to similar size. Adding the "+PiSA-SR" variants is also helpful. This is much better than just comparing against public checkpoints trained on different datasets. The training dataset could be useful if you release it.
  3. Fast inference: Single-step diffusion at 5.4 FPS is nice compared to other diffusion methods (0.5–2.5 FPS).
Critical weaknesses

W1. The ablation study does not clearly isolate the decomposition module's contribution.

Table 6 reports 19.84 dB for "w/o Adaptive Decomposition," compared with 21.81 dB for the full model. It is unclear whether this variant removes only D or also changes the encoders, fusion, and injection paths. Fig. 12 shows visual similarity between C_i and C_j, but does not quantify exposure invariance or establish an advantage over Gaussian/guided-filter and Retinex-based decompositions.

W2. The novelty appears incremental, and the contribution beyond existing designs is difficult to assess.

The components build on established techniques: paired decomposition training is related to RetinexNet (not cited), multi-layer conditioning is related to ControlNet/T2I-Adapter, SD-Turbo + LoRA follows existing diffusion adaptation approaches, and the data pipeline follows UNICE. Without a clearly specified direct-input baseline using the same pretrained backbone, the additional benefit of the E/C decomposition and gating remains unclear.

Questions for authors
  1. What exactly does "w/o Adaptive Decomposition" remove? Please describe which modules each ablation variant removes or retains, and clarify whether this variant provides a matched direct-input SD-Turbo baseline.
  2. Please compare with Gaussian/guided-filter and Retinex-based decompositions in the same pipeline, and report difference maps and quantitative cross-exposure consistency for ±3EV pairs.
  3. Is PiSA-SR trained on degraded inputs or baseline-enhanced outputs? Is one model shared across baselines, or is each trained separately? Please specify the inference settings and how it was adapted from super-resolution to same-resolution enhancement.
Reviewer_epKoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GICE, a diffusion-based framework for generalized image contrast enhancement under diverse illumination conditions and image degradations. The method introduces an exposure-aware decomposition module that separates exposure-dependent information from relatively stable image structures. A layer-adaptive conditioning mechanism then integrates these components into different layers of a pretrained diffusion model. The framework uses SD-Turbo for one-step image refinement and incorporates trainable VAE skip modules to preserve fine details. Experiments across low-light enhancement, exposure correction, and backlit image enhancement demonstrate the effectiveness of the proposed approach.

Critical strengths
  1. The paper considers a practically relevant problem. Real-world images often contain multiple types of contrast degradation, and developing a unified enhancement framework is more useful than optimizing separate models for individual degradation settings.

  2. The exposure-aware decomposition and layer-adaptive conditioning mechanisms provide a reasonable way to balance global contrast correction and local structure preservation. The use of a pretrained diffusion model also allows the framework to benefit from strong image priors.

  3. The experimental evaluation covers several enhancement tasks and includes both quantitative and qualitative comparisons. The one-step inference design is another practical advantage compared with diffusion approaches requiring multiple denoising steps.

Critical weaknesses
  1. The methodological contribution is relatively incremental. Image decomposition, adaptive feature conditioning, and diffusion-based restoration have been investigated in previous work. Although their combination is effective, the paper does not sufficiently demonstrate that the proposed design introduces a fundamentally new modeling capability.

  2. The distinction between exposure-variant and exposure-invariant components is not fully justified. Since illumination changes can also affect visible structures and textures, it is unclear whether the learned decomposition consistently corresponds to the intended semantic interpretation. More direct visualization and quantitative evaluation of the decomposition would strengthen the paper.

  3. The claimed generalization ability deserves further examination. The training procedure relies on multi-exposure data and synthetic degradation augmentation, which may not fully represent the complexity of real-world image corruption. Additional evaluations on challenging unseen degradation combinations would help establish the robustness of the proposed framework.

Questions for authors
  1. How does the proposed exposure decomposition compare with simpler feature disentanglement strategies under the same training conditions?

  2. What happens when the input contains degradation types that are not included in the augmentation pipeline?

  3. How much improvement is contributed by the proposed conditioning mechanism compared with the pretrained SD-Turbo backbone and trainable VAE skip connections?

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

Related papers

Discussion (0)

Sign in to comment.