Research map / Image generation / GANs
Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator
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 Direct Discriminative Optimization (DDO), which implicitly parameterizes a GAN-style discriminator via the likelihood ratio between a learnable model and a frozen reference, and iteratively finetunes pretrained diffusion and autoregressive image models to beat MLE limits.
Problem
Maximum likelihood training minimizes forward KL, causing mode-covering and limiting visual generation quality under limited capacity; GAN-style training needs extra discriminators and is unstable, and is costly for iterative-sampling models.
Approach
Parameterize the discriminator implicitly as sigmoid of the log-likelihood ratio between a learnable target model and a fixed reference model (DPO-style), so the GAN discriminator loss finetunes a likelihood-based model directly against real data and self-generated negatives, applied in multi-round self-play.
Key results
- DDO on EDM reduces CIFAR-10 unconditional FID from 1.97 to 1.38 (conditional 1.85 to 1.30) without guidance.
- EDM2-L + DDO reaches guidance-free FID 1.26 on ImageNet 512x512 (1.21 with autoguidance), and EDM2-S + DDO reaches 0.97 on ImageNet-64.
- VAR-d30 guidance-free FID drops from 4.74 to 1.79 on ImageNet 256x256 with DDO, and VAR-d16 from 11.33 to 3.12.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DDO to EDM | EDM · CIFAR-10 unconditional | FID: 1.97 -> 1.38 | Helped |
| Apply DDO to EDM | EDM · CIFAR-10 class-conditional | FID: 1.85 -> 1.30 | Helped |
| Apply DDO to EDM2-S | EDM2-S · ImageNet-64 class-conditional | FID: 1.60 -> 0.97 | Helped |
| Apply DDO to EDM2-L | EDM2-L · ImageNet 512×512 class-conditional, without guidance | FID: 1.96 -> 1.26 | Helped |
| Apply DDO to EDM2-L | EDM2-L · ImageNet 512×512 class-conditional, with guidance | FID: 1.77 -> 1.21 | Helped |
| Apply DDO and remove sampling tricks | VAR-d16 · ImageNet 256×256 class-conditional, guidance-free | FID: 11.33 -> 3.12 | Helped |
| Apply DDO and remove sampling tricks | VAR-d30 · ImageNet 256×256 class-conditional, guidance-free | FID: 4.74 -> 1.79 | Helped |
| Apply DDO and remove sampling tricks | VAR-d30 · ImageNet 256×256 class-conditional, with CFG | FID: 1.90 -> 1.73 | Helped |
| Continue training with the original diffusion loss (MLE) | Pretrained diffusion models · Class-conditional CIFAR-10 | FID | Hurt |
| Use a wide range of α, β settings (ablation) | DDO diffusion models · Class-conditional CIFAR-10 | FID | Helped |
| Use α values in [10,100] with β=0.02 (ablation) | VAR-d16 + DDO · ImageNet 256×256 class-conditional, with CFG | FID | Helped |
| Increase α (ablation) | VAR-d16 + DDO · ImageNet 256×256 class-conditional, guidance-free | FID | Helped |
Surprising findings
- Continued MLE training of the already-optimized EDM base model fails to improve and can degrade performance, while DDO consistently improves it.
- Guidance-free DDO-finetuned VAR-d16 and VAR-d30 outperform CFG-enhanced original VAR results while halving inference cost.
Limitations
- Requires hyperparameter sweeps over alpha and beta, with optimal values model-dependent.
- Theoretical divergence bounds rely on assumptions of bounded log-ratios, and the generalized objective with alpha, beta has no closed-form optimum.
- Evaluated only on class-conditional/unconditional image benchmarks; text-to-image not tested.
Experimental setup
Datasets: CIFAR-10 ImageNet
Benchmarks: CIFAR-10 ImageNet-64 ImageNet 512x512 ImageNet 256x256
Metrics: FID IS NFE
Baselines: EDM EDM2 VAR-d16 VAR-d30 Discriminator Guidance Autoguidance CCA StyleGAN-XL SiD2A DMD2
Builds on: Direct Preference Optimization GAN EDM EDM2 VAR Diffusion-DPO
Closest papers by meaning
- Dual Discriminator Generative Adversarial Nets (NeurIPS 2017 · similarity 0.90)
- Adaptive Density Estimation for Generative Models (NeurIPS 2019 · similarity 0.90)
- Ensembling Off-the-Shelf Models for GAN Training (CVPR 2022 · similarity 0.89)
- Refining Deep Generative Models via Discriminator Gradient Flow (ICLR 2021 · similarity 0.89)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.88)
- Discriminative Forests Improve Generative Diversity for Generative Adversarial Networks (AAAI 2024 · 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.