[Paper Review] GAN Dissection: Visualizing and Understanding Generative Adversarial Networks

David Bau, Jun-Yan Zhu+5 more|arXiv (Cornell University)|Nov 26, 2018

Cellular Automata and Applications217 citations

TL;DR

The paper introduces a framework to visualize and understand GANs by discovering interpretable units, testing their causal role via interventions, and applying this to debug and manipulate generated scenes.

ABSTRACT

Generative Adversarial Networks (GANs) have recently achieved impressive results for many real-world applications, and many GAN variants have emerged with improvements in sample quality and training stability. However, they have not been well visualized or understood. How does a GAN represent our visual world internally? What causes the artifacts in GAN results? How do architectural choices affect GAN learning? Answering such questions could enable us to develop new insights and better models. In this work, we present an analytic framework to visualize and understand GANs at the unit-, object-, and scene-level. We first identify a group of interpretable units that are closely related to object concepts using a segmentation-based network dissection method. Then, we quantify the causal effect of interpretable units by measuring the ability of interventions to control objects in the output. We examine the contextual relationship between these units and their surroundings by inserting the discovered object concepts into new images. We show several practical applications enabled by our framework, from comparing internal representations across different layers, models, and datasets, to improving GANs by locating and removing artifact-causing units, to interactively manipulating objects in a scene. We provide open source interpretation tools to help researchers and practitioners better understand their GAN models.

Motivation & Objective

• Identify units in GAN generators that correspond to object concepts (e.g., trees, tables).
• Quantify the causal influence of selected units on the presence or absence of objects in generated images.
• Explore how object concepts interact with context and background within scenes.
• Enable practical uses such as comparing representations across models, diagnosing artifacts, and interactive object manipulation.

Proposed method

• Dissection: measure agreement between unit activations and semantic segmentation masks to identify interpretable units (IoU-based).
• Intervention: ablate or insert sets of units and compute the average causal effect (ACE) on object presence using segmentation differences.
• Optimize a continuous intervention vector to efficiently select a subset of units that maximizes ACE (with L2 regularization).
• Compare representations across layers, models, and datasets to understand how architecture and training affect learned objects.
• Use artifact identification to ablate units and improve image quality (FID and human judgments).
• Provide open-source tools and demos for interactive exploration.

Experimental results

Research questions

• RQ1 Do GANs learn explicit, disentangled representations of object concepts within internal units?
• RQ2 Can we causally manipulate object presence in generated images by intervening on specific units?
• RQ3 How do layer depth, model variant, and training data affect the emergence of interpretable units?
• RQ4 Can we identify and remove artifact-causing units to improve realism?
• RQ5 How do context and surroundings influence the effect of inserting object concepts into scenes?

Key findings

• Some units emerge as object detectors (e.g., tables, sofas) with meaningful IoU matches to segmentation maps across diverse appearances.
• Mid-to-late layers tend to host interpretable units corresponding to objects and object parts, while early layers encode low-level features.
• Architectural choices (e.g., minibatch stddev, pixelwise normalization) affect the number and variety of interpretable units.
• Ablating artifact-causing units significantly improves image quality (lower FID and higher human preference) compared to random ablations.
• Interventions show that inserting or removing object units yields context-dependent results, revealing how GANs encode object-context relationships.
• Nullifying small, targeted sets of units can remove specific objects (e.g., windows, curtains) in some scenes, while others (e.g., tables) are harder to eliminate.