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. ADD uses 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 surpasses 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.

The fundamental challenge lies in that reducing sampling steps in diffusion models inevitably degrades output quality, as each step contributes to the gradual refinement of generated images. Prior distillation approaches such as progressive distillation and consistency distillation have made progress but still require 4-8 steps for acceptable quality. ADD overcomes this limitation by introducing a dual-objective training framework that combines the distributional knowledge of the teacher with the perceptual sharpness enforced by an adversarial critic, enabling unprecedented single-step generation quality.

Diffusion models have emerged as the dominant paradigm for high-quality image synthesis, powering applications from text-to-image generation to image editing and video synthesis. However, their main limitation is the requirement for many iterative denoising steps, typically 20-50, making them orders of magnitude slower than single-step generators like GANs. This computational burden prevents real-time applications and increases serving costs for commercial deployments. Distillation methods aim to compress the multi-step process into fewer steps, but existing approaches either sacrifice significant quality or still require 4-8 steps.

The speed limitation of diffusion models is not merely an engineering inconvenience but a fundamental barrier to an entire category of applications. Interactive image editing, real-time creative tools, and on-device generation all require sub-second inference times. While GANs achieve this speed, they lack the diversity and prompt-adherence of diffusion models. Latent Consistency Models (LCM) have demonstrated promising few-step generation but still produce noticeably blurry outputs at 1-2 steps. Our work addresses this gap by proposing ADD, which combines the best of both worlds: the distributional richness of diffusion models with the perceptual sharpness of adversarial training.

The key technical challenge in distilling diffusion models is the mode coverage vs. sample quality tradeoff. With fewer denoising steps, the model must traverse the same distribution space in larger jumps, which tends to average over nearby modes rather than sharply resolving individual modes. This manifests visually as blurriness — the hallmark failure mode of aggressive distillation. Previous approaches attempted to address this through improved training schedules, better loss weighting, and multi-scale supervision, but none achieved perceptual sharpness comparable to the original multi-step generation. ADD's adversarial component directly targets this specific failure mode by providing per-sample gradient signals that penalize blurriness.

Progressive distillation iteratively halves the number of sampling steps by training a student to match the output of two teacher steps in a single step. While effective at moderate step counts (4-8), it struggles in the extreme low-step regime because the accumulated approximation errors become visually apparent. Consistency models take a different approach by enforcing self-consistency along the probability flow ODE trajectory, enabling direct single-step generation. However, consistency distillation from powerful teachers like SDXL still produces outputs with noticeable artifacts and reduced sharpness. Score Distillation Sampling (SDS), originally proposed for 3D generation, provides a way to distill knowledge from a pretrained diffusion model but tends to produce over-smoothed results when used alone.

The use of adversarial training in conjunction with diffusion models is relatively unexplored. Prior work has used discriminators for improving sample quality in unconditional generation or for domain adaptation of diffusion models, but no prior work has combined adversarial objectives with score distillation for few-step sampling of foundation-scale models. Our insight is that these two signals are naturally complementary: score distillation captures the teacher's global distributional knowledge while the adversarial loss enforces local perceptual quality, addressing the specific failure mode of blurriness in low-step generation.

The connection between adversarial training and diffusion models can be understood through the lens of distribution matching at different granularities. Score distillation operates at the distribution level, pushing the student's output distribution toward the teacher's. The adversarial loss operates at the sample level, ensuring each individual output is realistic. In the low-step regime, distribution-level matching alone produces plausible but blurry outputs (averaging over modes), while sample-level discrimination alone may produce sharp but semantically inconsistent results. Their combination provides both distributional alignment and sample-level realism, addressing the complementary failure modes of each approach.

ADD combines two complementary training signals. The first is score distillation from a pretrained diffusion model teacher. Given a student-generated image, we add noise at a random timestep and use the teacher model to predict the denoised version. The difference between the teacher's prediction and the student's output provides a gradient signal that steers the student toward the distribution learned by the teacher. The second signal is an adversarial loss from a discriminator trained to distinguish real images from student-generated ones. While score distillation captures the global structure and semantics, the adversarial loss ensures local sharpness and perceptual quality, particularly in the extreme low-step regime where score distillation alone tends to produce blurry results.

The score distillation loss operates by adding noise to the student's output at a randomly sampled timestep t, then computing the denoising prediction of the frozen teacher model. Formally, the gradient is computed as the difference between the teacher's noise prediction and the added noise, weighted by a timestep-dependent scaling factor. This formulation allows the student to receive guidance from the teacher's learned distribution without requiring the teacher to generate full samples. The adversarial loss uses a discriminator that operates on features extracted from a pretrained vision model, providing rich perceptual features rather than raw pixel comparisons. We employ non-saturating GAN loss with R1 gradient penalty for stable training dynamics.

A critical design choice is the training strategy for supporting variable step counts. During training, we randomly sample the number of denoising steps (1, 2, or 4) and train the student to produce high-quality outputs for each. The noise schedule is carefully designed so that each step covers an equal portion of the total noise range, ensuring balanced contribution from each denoising step. We initialize the student from the pretrained SDXL weights and train with a combination of real images and teacher-generated images as discriminator references. The total training objective is a weighted sum of the score distillation loss and the adversarial loss, with the adversarial weight increasing for lower step counts to compensate for the greater blurriness in that regime.

We apply ADD to distill SDXL into a model we call SDXL Turbo. In single-step generation, SDXL Turbo achieves an FID of 28.3 on COCO-2017 30K, compared to 41.2 for StyleGAN-XL and 37.4 for LCM-LoRA. In four-step generation, SDXL Turbo achieves FID 23.4, matching SDXL's 50-step FID of 23.1. Human preference studies show that SDXL Turbo at 4 steps is preferred over SDXL at 50 steps in 48.2% of comparisons, indicating near-parity in perceived quality. The speedup is dramatic: single-step generation takes 0.34 seconds at 512x512 on a single A100 GPU, enabling real-time interactive applications for the first time with a foundation-scale diffusion model.

Ablation studies systematically validate each component's contribution. Removing the adversarial loss increases single-step FID from 28.3 to 38.9, confirming that adversarial training is essential for perceptual quality at low step counts. Removing score distillation degrades FID to 35.7 and produces images that lack the semantic richness and prompt adherence of the teacher. Using only progressive distillation yields FID 42.1 at single step, demonstrating that traditional distillation methods are insufficient in the extreme low-step regime. The CLIP score analysis further reveals that SDXL Turbo maintains comparable text-image alignment to SDXL across diverse prompt categories, indicating that the distillation process preserves the teacher's prompt-following capabilities.

Qualitative comparisons reveal distinctive characteristics of each step count. Single-step outputs exhibit slightly reduced fine detail but maintain strong compositional coherence, making them suitable for real-time previewing and iterative exploration. Two-step outputs recover most fine details while retaining near-instantaneous generation speed. Four-step outputs are visually indistinguishable from 50-step SDXL in most cases, representing the quality-speed sweet spot for production use. Failure cases primarily involve highly complex scenes with many small objects or text rendering, where the reduced step count limits the model's ability to iteratively refine fine spatial details.

We further investigate the discriminator architecture and its impact on generation quality. Our discriminator uses features from a pretrained DINOv2 ViT-L model, extracting intermediate representations at multiple layers to capture both low-level textures and high-level semantics. We compare this against discriminators using CLIP features, raw pixels, and learned CNN features. The DINOv2-based discriminator achieves the best FID across all step counts, with CLIP-based discriminator performing 2.1 FID points worse and pixel-based discriminator 5.8 points worse. This confirms that rich, self-supervised visual features are essential for the discriminator to provide meaningful gradients in the perceptual quality domain.

We have presented Adversarial Diffusion Distillation (ADD), the first approach to achieve high-fidelity, single-step image synthesis from foundation-scale diffusion models. By combining score distillation with adversarial training, ADD produces the SDXL Turbo model that matches SDXL quality in 4 steps and enables real-time generation in a single step. Our work opens the door to interactive, real-time creative applications powered by the full expressiveness of foundation diffusion models.

The broader implications of ADD extend beyond the specific SDXL Turbo model. The dual-objective framework of combining distributional distillation with adversarial refinement is applicable to any diffusion model architecture, suggesting a general recipe for accelerating foundation generative models. Future work may explore extending ADD to video diffusion models, 3D generation, and other modalities where the speed-quality tradeoff remains a critical bottleneck. The remaining gap between single-step and multi-step quality, particularly for complex scenes, suggests room for further improvement through better discriminator architectures or adaptive step allocation.