Adversarial Diffusion Distillation

By Axel Sauer, Dominik Lorenz, Andreas Blattmann, Robin Rombach. European Conference on Computer Vision, 2023

826 citations recorded by the source corpus.

We introduce Adversarial Diffusion Distillation (ADD), a novel training approach that efficiently samples large-scale foundational image diffusion models in just 1-4 steps while maintaining high image quality. We use score distillation to leverage large-scale off-the-shelf image diffusion models as a teacher signal in combination with an adversarial loss to ensure high image fidelity even in the low-step regime of one or two sampling steps. Our analyses show that our model clearly outperforms existing few-step methods (GANs, Latent Consistency Models) in a single step and reaches the performance of state-of-the-art diffusion models (SDXL) in only four steps. ADD is the first method to unlock single-step, real-time image synthesis with foundation models. Code and weights available under https://github.com/Stability-AI/generative-models and https://huggingface.co/stabilityai/ .

Subfields: Biophysics, Computer Vision and Pattern Recognition, Statistical and Nonlinear Physics

View Paper: https://www.semanticscholar.org/paper/d730d42bb655b3b44727d71c147f9758612043a8

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Biophysics

In Computer Vision and Pattern Recognition

In Statistical and Nonlinear Physics

Citations in the corpus

FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space (2025, positive): “FLUX.1 Kontext [pro] is trained with the flow objective followed by LADD [48].”
Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding (2024, methodology): “Algorithmic Acceleration Recently, various methods have been proposed to reduce the inference steps of diffusion-based text-to-image models [22, 31, 21, 37, 40].”
One-Step Diffusion with Distribution Matching Distillation (2023, methodology): “Our method is different from recent works that combine GANs with diffusion [68, 81–83], as our formulation is not grounded in GANs.”
Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (2024, positive): “This approach has also been applied to the distillation of diffusion models with adversarial objectives (Sauer et al., 2023b; 2024; Kang et al., 2024).”
Sora: A Review on Background, Technology, Limitations, and Opportunities of Large Vision Models (2024, methodology): “…visual encoders, such as the VAE encoder from Stable Diffusion [19], is advisable for most researchers while Sora ’s team is expected to train their own compression network with a decoder (the video generator) from scratch via the manner employed in training latent diffusion models [19, 35, 36].”
Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (2024, methodology): “Our largest models outperform state-of-the art open models such as SDXL (Podell et al., 2023), SDXL-Turbo (Sauer et al., 2023), Pixart-α (Chen et al., 2023), and closed-source models such as DALL-E 3 (Betker et al., 2023) both in quantitative evaluation (Ghosh et al., 2023) of prompt understanding a”
Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillation (2024, neutral): “The best results in the one-and few-step regime are currently achieved with methods that leverage adversarial training [50, 62, 49, 28], forcing the output distribution towards the real image manifold.”
DiffBIR: Towards Blind Image Restoration with Generative Diffusion Prior (2023, background): “There’re works [33, 44] that can already achieve satisfactory generation performance with only 1 ∼ 4 steps, thus the time-consuming problem can be solved in the future.”
Improved Distribution Matching Distillation for Fast Image Synthesis (2024, methodology): “Recent diffusion acceleration techniques have focused on speeding up the generation process through distillation [9, 10, 13–20, 22, 23, 30].”
PixArt-Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation (2024, background): “5 [40], Stable Turbo [42], Stable XL [37], Stable Cascade [36] and Playground-V2.”
Mean Flows for One-step Generative Modeling (2025, neutral): “One approach is to distill a pre-trained many-step diffusion model into a few-step model, e.g., [39, 13, 41] or score distillation [32, 50, 53].”
From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (2024, neutral): “Adversarial loss [18] is also used, sometimes in combination with other methods, to improve the quality of student output [29, 44, 70, 94, 99].”
One Step Diffusion via Shortcut Models (2024, neutral): “Alternatives to L2 distance have been proposed for distillation targets, such as adversarial (Sauer et al., 2023) or distribution-matching (Yin et al., 2024b;a) objectives.”