Denoising diffusion probabilistic models have revolutionized generative AI, producing high-fidelity images, videos, and structural designs. However, their exceptional quality comes from a fundamental computational bottleneck: they generate data by simulating a continuous trajectory through hundreds or thousands of sequential denoising steps. Taking a single giant step from pure noise to a clean image using a standard diffusion model introduces massive discretization errors, resulting in blurry, low-quality averages.
To unlock real-time applications—such as interactive image editing, high-frame-rate video generation, or rapid molecular design—researchers have developed techniques to collapse these multi-step trajectories into a single forward pass or a handful of discrete steps. This paradigm shift relies on three main insights: straightening the generation paths, matching probability distributions directly, and leveraging adversarial feedback.
Trajectory Straightening and Flow Matching
To understand why one-step generation is inherently difficult, consider the curved paths of standard diffusion ordinary differential equations (ODEs). The vector field changes dynamically at every time step, meaning a straight line drawn from a noise vector will completely miss the target data distribution.
One elegant solution is to mathematically "straighten" these transport paths during training. Cao et al. utilize a rectified flow paradigm combined with an analytical averaging technique over rotations (SO(3)-Averaged Flow). By straightening the generation path, the model can bypass the thousand-step stochastic simulations entirely, producing precise 3D molecular structures in a single step and speeding up inference by up to 50x.
Standard Diffusion Path (Curved) Rectified Flow Path (Straight)
[Noise] [Noise]
\ |
) <-- Requires many steps | <-- Single linear step
/ |
[Data] [Data]
Shifting the training objective from instantaneous velocity to average velocity provides another way to train one-step models from scratch. Geng introduces MeanFlows, deriving a mathematical identity that links these two quantities. By optimizing the model to predict the average velocity required to bridge the noise and data distributions, the framework eliminates the need for complex pre-training or curriculum learning, narrowing the performance gap between single-pass and multi-step models.
Distribution Matching and Score Distillation
Instead of training a model from scratch, an alternative approach distills a pre-trained, multi-step "teacher" diffusion model into a lightning-fast "student" network. The challenge is ensuring that the student's single-step outputs do not suffer from mode collapse or blurring.
To solve this, Yin et al. proposed Distribution Matching Distillation (DMD). Rather than forcing the student to mimic individual trajectories, DMD minimizes the Kullback-Leibler (KL) divergence between the student's output distribution and the teacher's distribution. It achieves this using two separate diffusion models that act as a system of gradients, providing "realness" and "fakeness" signals. This distribution-level alignment enables high-resolution image generation at 20 frames per second.
[Noise Vector] ---> ( Student Generator ) ---> [Generated Image]
|
+-------------------------------+
|
v
( Realness / Fakeness Gradients ) <-- [Teacher Distribution]
|
v
[Update Student Weights]
Distillation can also be achieved without relying on a massive dataset of real target images. Nguyen et al. introduced SwiftBrush, which repurposes Variational Score Distillation (VSD)—a technique originally designed for text-to-3D synthesis. By optimizing the student generator using the score functions of the teacher model, the authors compress the iterative denoising process into a single forward pass without requiring any training images, achieving a 20x inference speedup.
Adversarial Refinement and Discriminators
Pure distillation can sometimes hit a performance ceiling dictated by the teacher model's capabilities. To surpass this limit, researchers have re-introduced adversarial training principles, treating pre-trained diffusion networks or large-scale vision models as highly sophisticated critics.
Instead of relying on a standard teacher-student setup, Li et al. developed $D^3SR$ for real-world image super-resolution. This framework leverages a massive pre-trained model (SDXL) as a discriminator during training. By evaluating the structural realism of the single-step outputs against the discriminator's learned prior, the model bypasses traditional distillation ceilings, running significantly faster while utilizing fewer parameters than competing one-step models.
However, training single-step models with adversarial objectives can be notoriously unstable. Chen et al. address this in NitroFusion by employing a dynamic pool of constantly refreshing, specialized discriminators. This "panel of art critics" prevents the generator from overfitting to a single adversarial metric, ensuring that the single-step model produces vibrant, photorealistic images that match or exceed the aesthetic quality of multi-step models.
The Few-Step Sweet Spot
While compressing generation into exactly one step is the ultimate goal for real-time latency, forcing a model to map pure noise to complex data in a single pass can occasionally drop fine details. Consequently, a parallel branch of research targets a "few-step" sweet spot (typically 4 to 8 steps), which balances extreme speed with superior image fidelity.
Inference Steps vs. Image Quality Trade-off
FID Quality
^
| [Few-Step (4-8 steps)] <-- Optimal Balance
| *
| / \
| / \
| / * [Traditional Diffusion (100+ steps)]
| /
| * [One-Step]
+-------------------------------------------------------> Speed (FPS)
Heek et al. introduced Moment Matching Distillation, demonstrating that distilling a model into an 8-step regime allows the student to actually outperform its many-step teacher on standardized metrics like the Fréchet Inception Distance (FID).
To maximize the efficiency of these few-step models, the scheduling of time steps during training becomes critical. Mei et al. showed that using a consistent time sampling strategy during training—rather than randomizing time steps across a batch—significantly stabilizes conditional generation tasks. This approach allows models to perform complex image editing and super-resolution tasks with high fidelity in as few as 1 to 4 steps.
Key Foundational Work
Towards Performant and Reliable Undersampled MR Reconstruction via Diffusion Model Sampling (2022) — contributed 2 concept elements to this direction.