Research map / Image generation / GANs

Compressing Image-to-Image Translation GANs Using Local Density Structures on Their Learned Manifold

AAAI 2024 · Image generation › GANs · approach: GAN latent optimization and evaluation · 11 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Compresses Pix2Pix and CycleGAN generators and discriminators by pruning with a local-neighborhood density-matching objective on the generator's learned manifold, plus two collaborative pruning agents (GRU-based) that exchange architecture embeddings.

Problem

GAN compression for image-to-image translation: existing methods (KD, channel pruning, NAS) ignore the local density structure of the generator's learned manifold and often destabilize the generator-discriminator balance during pruning.

Approach

Partition the original generator's output manifold into local neighborhoods (sample plus k nearest neighbors found with a SwAV encoder fine-tuned on target images) and use an adversarial pruning objective so the pruned model preserves the density structure in each neighborhood; two pruning agents for G and D exchange architecture embeddings so capacities stay balanced.

Key results

• On Pix2Pix Cityscapes, MGGC reaches 83.60% MACs reduction and 44.63 mIoU vs. 42.71 for the original model.
• On CycleGAN Horse2Zebra, MGGC at 95.60% compression achieves 55.06 FID vs. 59.31 for GCC and 61.53 for the original model.
• Ablation without knowledge distillation still gives 58.64 FID on Horse2Zebra, outperforming the original model's 61.53.

Measured effects reported in the paper

Surprising findings

• Without knowledge distillation, the pruned CycleGAN on Horse2Zebra already outperforms the original model in FID.

Limitations

• Neighborhood search uses an approximate SwAV-based encoder rather than exact manifold neighbors.
• Pruning objective relies on a fixed pretrained generator and discriminator with frozen weights during pruning.

Experimental setup

Datasets: Cityscapes Edges2Shoes Horse2Zebra Summer2Winter

Benchmarks: Cityscapes Edges2Shoes Horse2Zebra Summer2Winter

Metrics: mIoU FID MACs

Baselines: GAN Compression CF-GAN CAT DMAD WKD RAKD Norm Pruning GCC CIL Pix2Pix 0.5x Co-Evolution GAN Slimming

Builds on: GCC Pix2Pix CycleGAN SwAV GAN Compression

Closest papers by meaning

• Content-Aware GAN Compression (CVPR 2021 · similarity 0.91)
• GAN Compression: Efficient Architectures for Interactive Conditional GANs (CVPR 2020 · similarity 0.91)
• RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.90)
• TraVeLGAN: Image-To-Image Translation by Transformation Vector Learning (CVPR 2019 · similarity 0.90)
• Singular Value Scaling: Efficient Generative Model Compression via Pruned Weights Refinement (AAAI 2025 · similarity 0.88)
• E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation (ICML 2024 · similarity 0.88)