Research map / Image generation / GANs
Consistency-GAN: Training GANs with Consistency Model
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.
Proposes Consistency-GAN, a GAN training scheme that injects instance noise via a consistency-model-based mapping module (few steps) and uses a step-conditional discriminator to get quality, diversity and fast sampling together.
Problem
GANs suffer from limited sample diversity (mode collapse) and training instability, while diffusion models have high sample quality and diversity but very slow iterative sampling; the paper targets all three criteria (quality, diversity, speed) at once.
Approach
Trains a consistency mapping module that maps points on a probability-flow ODE trajectory back to their origin in a few steps, and uses it to inject instance noise into real and generated samples during adversarial training, with a discriminator conditioned on the mapping step n.
Key results
- On CIFAR-10, Consistency-GAN reaches FID 3.16, IS 9.76 and recall 0.47, versus Projected GAN FID 3.39 and Diffusion-GAN FID 4.92.
- With NFE of 5 on CIFAR-10, it beats DDPM (NFE 1000, FID 3.21) and DDIM (NFE 10, FID 8.23) on IS and FID.
- Reaching FID 5 takes Diffusion-GAN 20,000 iterations (over 60 hours) versus about 3,000 steps (around 4 hours) for Consistency-GAN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Consistency-GAN instead of Projected GAN | CIFAR-10 image generation · CIFAR-10 | FID: 3.39 -> 3.16 | Helped |
| Use Consistency-GAN instead of BigGAN | CIFAR-10 image generation · CIFAR-10 | Recall: 0.36 -> 0.47 | Helped |
| Use Consistency-GAN instead of DDPM | CIFAR-10 image generation · CIFAR-10 | NFE: 1000 -> 5 | Helped |
| Use Consistency-GAN instead of Diffusion-GAN | CIFAR-10 image generation · CIFAR-10 | FID: 4.92 -> 3.16 | Helped |
| Use Consistency-GAN instead of Diffusion-GAN | STL-10 image generation · STL-10 | FID: 6.91 -> 6.8 | Helped |
| Use Consistency-GAN instead of Diffusion-GAN | CelebA 256×256 image generation · CelebA 256×256 | FID: 4.75 -> 3.93 | Helped |
| Inject instance noise through the consistency mapping module during adversarial training (ablation) | Consistency-GAN · CIFAR-10 | training stability | Helped |
| Set the consistency-mapping learning rate to 0.02 (ablation) | Consistency-GAN · CIFAR-10 | test accuracy | Helped |
| Use L2 rather than L1 or LPIPS as the consistency-mapping distance metric (ablation) | Consistency-GAN · CIFAR-10 | convergence speed | Helped |
| Increase the number of mapping steps up to T=5 (ablation) | Consistency-GAN · CIFAR-10 | FID | Helped |
| Increase the number of mapping steps beyond a certain threshold (ablation) | Consistency-GAN · CIFAR-10 | Recall | Hurt |
Surprising findings
- Noise injection slows discriminator convergence slightly but stabilizes training, per loss curves compared against a no-noise run.
Limitations
- Comparisons with diffusion models rely on reported or single-setting baselines; limited ablation scope beyond CIFAR-10 hyperparameters.
- Consistency mapping training is a separate stage before adversarial training, adding pipeline complexity.
Experimental setup
Datasets: CIFAR-10 STL-10 CelebA
Benchmarks: CIFAR-10 STL-10 CelebA
Metrics: Inception Score FID Recall NFE
Baselines: BigGAN StyleGAN2 TransGAN Projected GAN DDPM DDIM Diffusion-GAN
Builds on: Diffusion-GAN consistency models Projected GAN DDIM
Closest papers by meaning
- Improved Consistency Regularization for GANs (AAAI 2021 · similarity 0.91)
- Consistency Regularization for Generative Adversarial Networks (ICLR 2020 · similarity 0.91)
- Improved Techniques for Training Consistency Models (ICLR 2024 · similarity 0.91)
- Enhancing Consistency-Based Image Generation via Adversarialy-Trained Classification and Energy-Based Discrimination (NeurIPS 2024 · similarity 0.90)
- ACT-Diffusion: Efficient Adversarial Consistency Training for One-step Diffusion Models (CVPR 2024 · similarity 0.88)
- Improved Diffusion-based Generative Model with Better Adversarial Robustness (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.