Research map / Image generation / GANs
Adversarial Score Distillation: When score distillation meets GAN
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.
Revisits Score Distillation Sampling as a WGAN generator loss with a pretrained diffusion model as implicit discriminator, and proposes Adversarial Score Distillation (ASD), which keeps the discriminator optimizable and trains it with the full WGAN discriminator objective, improving CFG-scale robustness for 2D distillation, text-to-3D, and image editing.
Problem
Score distillation methods (SDS, VSD) are sensitive to the classifier-free guidance scale, yielding over-smoothing or instability at small scales and over-saturation at large scales, with unclear methodological origins.
Approach
Reinterprets SDS as the generator loss of a WGAN whose discriminator is a diffusion model with a prompt embedding/LoRA for fake samples; SDS fixes this discriminator and VSD optimizes it incompletely, so ASD optimizes the discriminator with the complete WGAN discriminator loss (including a penalty term with hyperparameter gamma).
Key results
- On text-to-3D, ASD achieves CLIP B/32 score 0.3314 versus 0.3282 for SDS and 0.3203 for VSD.
- In a user study against VSD, ASD was preferred for overall quality 65.5% vs 34.5%.
- ASD recovers from over-saturated initialization where SDS with huge CFG remains over-saturated (qualitative).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ASD instead of ProlificDreamer (VSD) | Text-to-3D · Text-to-3D | CLIP B/32: 0.3203 -> 0.3314 | Helped |
| Use ASD instead of ProlificDreamer (VSD) | Text-to-3D · Text-to-3D | CLIP B/16: 0.3258 -> 0.3376 | Helped |
| Use ASD instead of ProlificDreamer (VSD) | Text-to-3D · Text-to-3D | CLIP L/14: 0.2789 -> 0.2854 | Helped |
| Use ASD instead of ProlificDreamer (VSD) | Text-to-3D user study · Text-to-3D user study | Details: 35.3 -> 64.7 | Helped |
| Use ASD instead of ProlificDreamer (VSD) | Text-to-3D user study · Text-to-3D user study | Semantic Alignment: 45.7 -> 54.3 | Helped |
| Use ASD instead of ProlificDreamer (VSD) | Text-to-3D user study · Text-to-3D user study | Overall Quality: 34.5 -> 65.5 | Helped |
| Increase LoRA updates to 50 steps per iteration (ablation) | VSD 2D score distillation · 2D score distillation | image quality | Hurt |
| Use only the classifier term (ablation) | Score distillation · 2D and 3D distillation | image saturation | Hurt |
| Set γ to 0, reducing discriminator optimization to VSD (ablation) | ASD · Score distillation | distillation stability | Hurt |
| Share the same added noise between generator and discriminator optimization (ablation) | ASD · 3D NeRF distillation | distillation speed: 30 ∼ 50% | Helped |
Surprising findings
- Only using the classifier term of the SDS gradient behaves like SDS with a very large CFG scale and tends to over-saturate in both 2D and 3D.
- Updating the VSD LoRA branch with 50 steps per iteration degrades results to over-smoothed images similar to SDS, consistent with WGAN failing when approximating the exact Wasserstein distance.
Limitations
- ASD speed is similar to VSD; distilling a 3D NeRF takes more than 5 hours on a single A800 GPU.
- gamma is set by fixed per-task values (-1 for 2D, -0.5 for text-to-3D) rather than dynamically tuned.
- Both ASD and VSD eventually change objects or structures due to their adversarial nature.
Experimental setup
Benchmarks: CLIP score (text-to-3D renders)
Metrics: CLIP B/32 CLIP B/16 CLIP L/14 user study preference
Baselines: SDS (DreamFusion) VSD (ProlificDreamer) DDS Classifier Score Distillation
Builds on: Score Distillation Sampling Variational Score Distillation Wasserstein GAN Delta Denoising Score ProlificDreamer
Closest papers by meaning
- SMaRt: Improving GANs with Score Matching Regularity (ICML 2024 · similarity 0.88)
- Improving the Training of Data-Efficient GANs via Quality Aware Dynamic Discriminator Rejection Sampling (CVPR 2025 · similarity 0.87)
- Indirect Gradient Matching for Adversarial Robust Distillation (ICLR 2025 · similarity 0.87)
- Adversarial Score identity Distillation: Rapidly Surpassing the Teacher in One Step (ICLR 2025 · similarity 0.87)
- Towards Adversarially Robust Dataset Distillation by Curvature Regularization (AAAI 2025 · similarity 0.86)
- Robust Distillation via Untargeted and Targeted Intermediate Adversarial Samples (CVPR 2024 · similarity 0.86)
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.