[Paper Review] RelGAN: Multi-Domain Image-to-Image Translation via Relative Attributes

Po-Wei Wu, Yujing Lin+3 more|arXiv (Cornell University)|Aug 20, 2019

Generative Adversarial Networks and Image SynthesisComputer Science29 references48 citations

TL;DR

RelGAN introduces relative attributes for multi-domain image translation, enabling continuous, targeted edits without specifying all attributes and providing improved realism and interpolation over prior target-attribute methods.

ABSTRACT

Multi-domain image-to-image translation has gained increasing attention recently. Previous methods take an image and some target attributes as inputs and generate an output image with the desired attributes. However, such methods have two limitations. First, these methods assume binary-valued attributes and thus cannot yield satisfactory results for fine-grained control. Second, these methods require specifying the entire set of target attributes, even if most of the attributes would not be changed. To address these limitations, we propose RelGAN, a new method for multi-domain image-to-image translation. The key idea is to use relative attributes, which describes the desired change on selected attributes. Our method is capable of modifying images by changing particular attributes of interest in a continuous manner while preserving the other attributes. Experimental results demonstrate both the quantitative and qualitative effectiveness of our method on the tasks of facial attribute transfer and interpolation.

Motivation & Objective

• Address limitations of binary, target-attribute based multi-domain translation by using relative attribute changes.
• Enable continuous, fine-grained control over attribute edits while preserving non-target attributes.
• Improve interpolation quality between original and edited images through dedicated discriminators and loss terms.
• Demonstrate effectiveness on facial attribute transfer, reconstruction, and interpolation across multiple high-quality datasets.

Proposed method

• Represent domains with an n-dimensional attribute vector a, and define relative attributes v = âˆhat a - a to specify the desired change.
• Use a single generator G conditioned on (x, v) with three discriminators: Real (unconditional realism), Match (whether (x, v, x') matches the translation), and Interp (predict degree of interpolation).
• Train with adversarial losses for realism (Real), conditional matching (Match) using real and wrong triplets, and an interpolation loss (Interp) to regularize smooth attribute changes.
• Impose reconstruction regularizers: cycle-reconstruction L1 loss between x and G(G(x, v), -v) and self-reconstruction loss when v = 0 to preserve identity and background details.
• Incorporate an interpolation discriminator that predicts the interpolation degree α for G(x, αv) to encourage smooth, realistic transitions.
• Adopt an orthogonal regularization term and train with LSGAN-GP stabilizers; use switchable normalization in the generator.

Experimental results

Research questions

• RQ1 Can relative attributes provide finer-grained, continuous attribute control compared to binary target attributes in multi-domain image translation?
• RQ2 How can the model selectively modify only the attributes of interest while preserving unchanged attributes and overall identity?
• RQ3 Does adding an interpolation discriminator improve the quality and smoothness of attribute interpolation?
• RQ4 What are the empirical benefits of RelGAN on facial attribute transfer, reconstruction, and interpolation across diverse datasets?

Key findings

• RelGAN achieves the lowest Fréchet Inception Distance (FID) compared with StarGAN and AttGAN across CelebA, CelebA-HQ, and FFHQ settings, indicating higher visual quality.
• Classification accuracy on generated images is highest for RelGAN in multiple attributes, indicating better fidelity of translated attributes.
• RelGAN preserves unchanged attributes more effectively than prior methods, and demonstrates smoother, more realistic interpolation between original and edited images.
• Ablation studies show that full loss (Real + Match + Cycle/Self + Interp with orthogonal regularization) yields the best reconstruction and interpolation results.
• User studies favor RelGAN in most attribute transfer and reconstruction tasks, with overall preference across multiple tasks.