Research map / Image generation / Fast diffusion sampling
Adversarial Diffusion Distillation
Axel Sauer, Dominik Lorenz, Andreas Blattmann, Robin Rombach
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 |
|---|---|---|---|
| add image conditioning to the discriminator (in addition to or instead of text) (ablation) | ADD-M discriminator · COCO2017 5k (FID, CLIP score) | FID / CLIP score | Improved |
| use text conditioning on the discriminator (ablation) | ADD-M discriminator · COCO2017 5k | FID / CLIP score | Improved |
| use DINOv2 ViT as discriminator feature network (ablation) | ADD-M · COCO2017 5k | FID / CLIP score | Improved |
| initialize the student from a pretrained diffusion model instead of from scratch (ablation) | ADD-student · COCO2017 5k | FID / CLIP score | Improved |
| add adversarial loss to the distillation loss (ablation) | ADD · COCO2017 5k | FID / CLIP score | Improved |
| use SDS or NFSD distillation weighting instead of exponential weighting (ablation) | ADD distillation loss · COCO2017 5k | FID / CLIP score | Mixed |
| use a bigger student and teacher (SDXL instead of SD) (ablation) | ADD · COCO2017 5k | FID / CLIP score | No clear effect |
| take several consecutive steps with the teacher (ablation) | ADD distillation · COCO2017 5k | FID / CLIP score | No clear effect |
| distill SDXL with ADD and sample in four steps instead of 50-step SDXL | ADD-XL · PartiPrompts user study (100 prompts) | human preference / ELO | Improved |
| single-step ADD-XL versus four-step LCM-XL | ADD-XL · PartiPrompts user study | human preference / ELO | Improved |
| distill into student with adversarial loss | ADD-XL versus teacher SDXL · sample diversity | diversity | Hurt |
Closest papers by meaning
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.92)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.91)
- MimicDiffusion: Purifying Adversarial Perturbation via Mimicking Clean Diffusion Model (CVPR 2024 · similarity 0.91)
- Structure-Guided Adversarial Training of Diffusion Models (CVPR 2024 · similarity 0.90)
- Diffusion Adversarial Post-Training for One-Step Video Generation (ICML 2025 · similarity 0.90)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.90)
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.
