WaveFill: A Wavelet-based Generation Network for Image Inpainting

• Large corrupted regions with complex texture and structural patterns make realistic, high-fidelity inpainting challenging.Source: §1 · p. 1
• Reconstruction losses favor low-frequency global structure, while adversarial losses favor high-frequency texture, and jointly optimizing them in spatial features creates inter-frequency conflicts.Source: §1 · p. 1
• Weighted losses in GMCNN [35] and Coarse-to-Fine refinement in GC [40] still leave frequency conflicts or feature-alignment inconsistencies.GC's refinement network retains inter-frequency conflicts, while its two-stage features can be inconsistent across structure and texture.Source: §1 · p. 2
• WaveFill uses DWT to separate frequency bands, applies frequency-appropriate losses, and uses FRAN to align and fuse their features.The completed bands are transformed back to the spatial domain with IDWT.Source: §1 · p. 2
• Extensive experiments over multiple datasets show that WaveFill achieves superior inpainting compared with state-of-the-art methods.Source: §1 · p. 2

2.1. Image Inpainting

• Diffusion methods can fail to recover meaningful structures with little global information, while patch-based methods struggle with semantics in non-stationary data.Patch-based methods work well for stationary texture.Source: §2.1 · p. 2
• Deep inpainting methods include GAN completion [28], Coarse-to-Fine recovery [39], edge prediction [26], multi-branch receptive fields [35], deep/shallow feature representations [21], and partial or gated convolutions [20] [40].Source: §2.1 · p. 2
• Most prior methods work in the spatial domain, where mixed frequency information can introduce inter-frequency conflicts; WaveFill instead applies different objectives to separate frequency bands.Source: §2.1 · p. 2

2.2. Wavelet-based Methods

• Earlier wavelet inpainting combined wavelets with diffusion, variational, or exemplar-based methods, but their hand-crafted features could not generate meaningful content for large corrupted regions.The cited examples include Chan et al., Dobrosotskaya et al., and Zhang and Dai.Source: §2.2 · p. 2
• Unlike deep wavelet methods that concatenate frequency bands for convolution, WaveFill uses separate branches and incorporates features from other branches for completion.Source: §2.2 · p. 3

3.1. Overview

• Three branches complete LowFreq, Lv2-HighFreq, and Lv1-HighFreq separately, then inverse DWT returns their predictions to the spatial domain.L1 reconstruction supervises LowFreq, while adversarial losses supervise both high-frequency branches.Source: §3.1 · p. 3

3.2. Wavelet Decomposition

• The 2D DWT iteratively decomposes LL into 3Nw + 1 sub-bands; with Haar filters and Nw = 2, the network uses one low- and two high-frequency inputs.The resulting inputs are LL2, concatenated detail bands at level 2, and concatenated detail bands at level 1.Source: §3.2 · p. 3

3.3. Frequency Region Attentive Normalization

• FRAN uses low-frequency self-attention to aggregate high-frequency features, then derives modulation parameters that inject the aligned information into normalized low-frequency features.This feature alignment addresses the different statistics of the bands and the risk of cross-band inconsistencies that can produce blur and artifacts.Source: §3.3 · p. 4, Figure 3 · p. 4

3.4. Network Architecture

• The generator has separate low- and high-frequency branches, with high-frequency branches using FRAN ResBlocks; two discriminators target the two high-frequency levels.The Lv1-HighFreq branch also combines aligned features from the preceding branches, and differentiable DWT and IDWT allow end-to-end training.Source: §3.4 · pp. 4–5

3.5. Loss Functions

• Training uses low-frequency wavelet-domain L1 reconstruction, hinge adversarial losses for both high-frequency-branch discriminators, feature matching on both discriminators to stabilize training, and VGG-19 perceptual loss for perceptual and semantic discrepancy.Source: §3.5 · p. 5

4.1. Experimental Settings

• Experiments use three public datasets, compare against established inpainting methods, and evaluate perceptual quality, reconstruction error, PSNR, and SSIM.The setup includes central-square and irregular masks; the full model runs at 0.138 seconds per 256 × 256 image on a single GPU.Source: §4.1 · pp. 5–6, §4.2 · p. 6

4.2. Quantitative Evaluation

• WaveFill outperforms existing methods across evaluation metrics on CelebA-HQ square masks and across mask ratios on irregular-mask experiments.The comparisons cover CelebA-HQ and Places2; Paris StreetView quantitative results are reported in supplementary materials.Source: Table 1 · p. 6, Table 2 · p. 7, §4.2 · p. 7

4.3. Qualitative Evaluations

• On CelebA-HQ and Places2, WaveFill produces more semantically plausible inpainting with fewer artifacts and more texture detail than the compared methods.GMCNN and EC show semantic or edge-region inconsistencies on CelebA-HQ, while GC shows artifacts and blurry textures; Places2 comparisons also note distorted structures and limited semantic plausibility in competing outputs.Source: Figure 4 · p. 6, §4.3 · p. 7, Figure 5 · p. 7

4.4. User Study

• In a user study comprising 75 multiple-choice questions, 20 volunteers with image-processing backgrounds voted on realism, and WaveFill outperformed state-of-the-art methods by large margins.Source: §4.4 · p. 8, Figure 6 · p. 8

4.5. Ablation Study

• Wavelet transformation outperforms DCT, while FRAN addresses cross-frequency feature alignment that concatenation and SPADE do not provide.The ablations report that DCT degrades inpainting due to lost spatial information; the study evaluates Paris StreetView irregular masks.Source: §4.5 · p. 8, Table 3 · p. 8, Figure 7 · p. 8

5. Conclusion

• WaveFill separates low- and high-frequency information in the wavelet domain and fills corrupted regions explicitly.The framework processes the frequency bands separately.Source: §5 · p. 8
• FRAN aligns and fuses multi-frequency features, especially those within missing regions, to ensure inpainting consistency across bands.Source: §5 · p. 8
• Experiments show superior image inpainting for both rectangle and free-form masks.Source: §5 · p. 8
• Future work will study adapting wavelet decomposition and separate frequency-band processing to other image recovery and generation tasks.Source: §5 · p. 8