Research map / Image generation / GANs
Unifying GANs and Score-Based Diffusion as Generative Particle Models
Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickael Chen, Alain Rakotomamonjy
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace EDM (generator-free diffusion) with Discriminator Flow | particle generative model · MNIST | FID: 3 -> 4 | No clear effect |
| Replace EDM with Discriminator Flow | particle generative model · CelebA 64x64 | FID: 10 -> 41 | Hurt |
| Replace GAN with Score GAN (score-based generator training) | generator-based model · MNIST | FID: 3 -> 15 | Hurt |
| Replace GAN with Score GAN (score-based generator training) | generator-based model · CelebA 64x64 | FID: 19 -> 35 | Hurt |
| Add a generator to diffusion (Score GAN instead of EDM) for one-step sampling | EDM · MNIST | sampling cost / FID: 3 -> 15 | Mixed |
| Remove generator from GAN (Discriminator Flow instead of GAN) on MNIST | GAN · MNIST | FID: 3 -> 4 | No clear effect |
| Remove generator from GAN (Discriminator Flow instead of GAN) on CelebA | GAN · CelebA 64x64 | FID: 19 -> 41 | Hurt |
| Use Discriminator Flow instead of EDM for early stopping of generation | EDM · MNIST | convergence speed to data distribution | Improved |
| Use Discriminator Flow with fewer steps vs EDM Euler sampler | Discriminator Flow · MNIST | FID vs NFE | Hurt |
| Vary number of score update steps K from 1 to 10 in Score GAN (ablation) | Score GAN · MNIST | FID | No clear effect |
| Add stochastic component (Brownian noise) to Score GAN updates (ablation) | Score GAN · image data | FID | Hurt |
| Use EDM parameters (mu_data=2, mu_rho=1) in Score GAN instead of NCSN parameters (ablation) | Score GAN · MNIST | mode collapse | Hurt |
Closest papers by meaning
- Rectifying GAN Gradient Supervision Via Score Functions (TPAMI 2026 · similarity 0.90)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.90)
- Generative Adversarial Diffusion (ICCV 2025 · similarity 0.89)
- Diffusion-GAN: Training GANs with Diffusion (ICLR 2023 · similarity 0.88)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.88)
- Deep MMD Gradient Flow without adversarial training (ICLR 2025 · similarity 0.87)
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.
