Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Data relativistic uncertainty framework for low-illumination anime scenery image enhancement

Yiquan Gao, John See

TL;DR

This paper tackles the domain gap and data scarcity in low-illumination anime scenery image enhancement by building the first unpaired anime dataset and introducing a Data Relativistic Uncertainty (DRU) framework. DRU quantifies illumination uncertainty and reweights training losses to emphasize more confident illumination samples, integrated with EnlightenGAN to improve perceptual and aesthetic quality. Empirical results show that DRU-EnlightenGAN variants outperform state-of-the-art unsupervised LLIE methods, with Transformer-based backbones (e.g., ViT-B16) delivering the strongest overall performance and robustness to data noise. The proposed data-centric learning paradigm offers a path toward robust anime LIE and may generalize to broader vision and language tasks.

Abstract

By contrast with the prevailing works of low-light enhancement in natural images and videos, this study copes with the low-illumination quality degradation in anime scenery images to bridge the domain gap. For such an underexplored enhancement task, we first curate images from various sources and construct an unpaired anime scenery dataset with diverse environments and illumination conditions to address the data scarcity. To exploit the power of uncertainty information inherent with the diverse illumination conditions, we propose a Data Relativistic Uncertainty (DRU) framework, motivated by the idea from Relativistic GAN. By analogy with the wave-particle duality of light, our framework interpretably defines and quantifies the illumination uncertainty of dark/bright samples, which is leveraged to dynamically adjust the objective functions to recalibrate the model learning under data uncertainty. Extensive experiments demonstrate the effectiveness of DRU framework by training several versions of EnlightenGANs, yielding superior perceptual and aesthetic qualities beyond the state-of-the-art methods that are incapable of learning from data uncertainty perspective. We hope our framework can expose a novel paradigm of data-centric learning for potential visual and language domains. Code is available.

Data relativistic uncertainty framework for low-illumination anime scenery image enhancement

TL;DR

This paper tackles the domain gap and data scarcity in low-illumination anime scenery image enhancement by building the first unpaired anime dataset and introducing a Data Relativistic Uncertainty (DRU) framework. DRU quantifies illumination uncertainty and reweights training losses to emphasize more confident illumination samples, integrated with EnlightenGAN to improve perceptual and aesthetic quality. Empirical results show that DRU-EnlightenGAN variants outperform state-of-the-art unsupervised LLIE methods, with Transformer-based backbones (e.g., ViT-B16) delivering the strongest overall performance and robustness to data noise. The proposed data-centric learning paradigm offers a path toward robust anime LIE and may generalize to broader vision and language tasks.

Abstract

By contrast with the prevailing works of low-light enhancement in natural images and videos, this study copes with the low-illumination quality degradation in anime scenery images to bridge the domain gap. For such an underexplored enhancement task, we first curate images from various sources and construct an unpaired anime scenery dataset with diverse environments and illumination conditions to address the data scarcity. To exploit the power of uncertainty information inherent with the diverse illumination conditions, we propose a Data Relativistic Uncertainty (DRU) framework, motivated by the idea from Relativistic GAN. By analogy with the wave-particle duality of light, our framework interpretably defines and quantifies the illumination uncertainty of dark/bright samples, which is leveraged to dynamically adjust the objective functions to recalibrate the model learning under data uncertainty. Extensive experiments demonstrate the effectiveness of DRU framework by training several versions of EnlightenGANs, yielding superior perceptual and aesthetic qualities beyond the state-of-the-art methods that are incapable of learning from data uncertainty perspective. We hope our framework can expose a novel paradigm of data-centric learning for potential visual and language domains. Code is available.
Paper Structure (13 sections, 9 equations, 4 figures, 3 tables)

This paper contains 13 sections, 9 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Low-light Datasets
  4. Unsupervised and Zero-Reference Enhancement
  5. Illumination Data Uncertainty
  6. Methodology
  7. Unpaired Anime Data Construction Method
  8. Data Relativistic Uncertainty Framework
  9. Experiments
  10. Implementation Details
  11. State-of-the-Art Comparison
  12. Ablation Study
  13. Conclusion

Figures (4)

  • Figure 1: Domain gap between natural and anime image enhancement models: EnlightenGAN trained on natural data is prone to color bias and artifacts, such as the blue masks on enhanced results, while our model restores more aesthetically pleasing anime contents.
  • Figure 2: Overview of our data construction method for unpaired anime scenery dataset.
  • Figure 3: Visualize how illumination uncertainty of dark and bright images passes through each component in EnlightenGAN and finally affects the various loss terms of EnlightenGAN.
  • Figure 4: Qualitative comparisons among different methods. Zoom in for detailed view.