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/ .
核心贡献 · Key contributions
提出对抗扩散蒸馏（ADD），一种仅需 1-4 步即可采样基础图像扩散模型的方法。 Introduces Adversarial Diffusion Distillation (ADD), a method for 1-4 step sampling of foundational image diffusion models.
结合对抗损失和分数蒸馏，在少步生成中实现高保真度。 Combines adversarial loss and score distillation to achieve high fidelity in few-step generation.
根据用户研究，ADD-XL 在 4 步时超越教师模型 SDXL，在 1 步时超越 LCM-XL。 ADD-XL outperforms teacher SDXL in 4 steps and LCM-XL in 1 step, per user studies.
首个实现基础模型单步实时图像合成的方法。 First method to enable single-step real-time image synthesis with foundation models.
消融研究表明预训练学生模型、判别器条件设置和损失组合的重要性。 Ablation studies show importance of pretrained student, discriminator conditioning, and loss combination.
保留迭代优化能力，不同于纯 GAN 方法。 Retains iterative refinement capability, unlike pure GAN approaches.
局限 · Limitations
需要预训练扩散模型作为教师和学生初始化。 Requires pretrained diffusion model as teacher and student initialization.
对抗训练可能引入不稳定性，需要仔细调参。 Adversarial training may introduce instability and requires careful tuning.
增强的真实感以略微降低样本多样性为代价。 Enhanced realism at cost of slightly decreased sample diversity.
评估限于 512x512 分辨率；未测试更高分辨率。 Evaluation limited to 512x512 resolution; higher resolutions not tested.
性能依赖于蒸馏损失的权重调度选择。 Performance depends on choice of weighting schedule for distillation loss.
论文章节 · Sections（共 13）
摘要Abstract
1 引言1 Introduction
2 背景2 Background
3 方法3 Method
3.1 训练流程3.1 Training Procedure
3.2 对抗损失3.2 Adversarial Loss
3.3 分数蒸馏损失3.3 Score Distillation Loss
4 实验4 Experiments
4.1 消融研究4.1 Ablation Study
4.2 与最新技术的定量比较4.2 Quantitative Comparison to State-of-the-Art