Accelerating Diffusion Sampling via Distillation and Trajectory Rectification

This direction is split into other directions: iterative sampling latency (iterative sampling overhead). Its last version stays readable and citable here.

Version 1 of 1164 papers165 conceptsfrozen 2026-10-04

Standard diffusion models generate high-fidelity images by reversing a gradual noise process. However, this reversal requires solving a complex Ordinary Differential Equation (ODE) through hundreds or thousands of iterative denoising steps. This iterative nature creates a significant computational bottleneck, preventing real-time applications in fields ranging from interactive art to clinical medical imaging.

To bridge this gap, researchers are developing techniques to "short-circuit" the denoising process, reducing the required steps from hundreds to as few as one to four without sacrificing visual quality.

Straightening the Probability Flow

The primary reason diffusion models are slow is that the "path" from pure noise to a clean image is often curved and complex. Solving a curved path requires many small, incremental steps to avoid drifting off-course.

Xingchao Liu and colleagues introduced InstaFlow InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation, which utilizes "Rectified Flow" to straighten these trajectories. By making the mapping between noise and images a straight line, the model can leap from start to finish in a single inference step. Similarly, Fu-Yun Wang and the team behind Rectified Diffusion Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow argue that the true power of this approach lies in retraining on matched noise-image pairs, which simplifies the learning objective and accelerates convergence.

Mapping Consistency Across Time

While rectification straightens the path, Consistency Models (CMs) aim to eliminate the path entirely. The core intuition is that any point along a specific denoising trajectory should ideally map back to the same starting image.

Alexander William Bergman and Fu-Yun Wang expanded this with Phased Consistency Models (PCMs) Phased Consistency Models, which break the trajectory into manageable sub-segments. This allows the model to maintain high controllability and detail even in 1–4 steps. To add fine-grained control (like poses or edges) to these fast models, Jie Xiao et al. developed CCM CCM: Adding Conditional Controls to Text-to-Image Consistency Models, proving that lightweight adapters can bring the flexibility of ControlNet to high-speed consistency frameworks.

Distillation from Teacher to Student

Another dominant strategy is distillation, where a "student" model learns to mimic the output of a pre-trained, multi-step "teacher" model.

Optimization and Reinforcement Learning

Beyond mimicking trajectories, some researchers treat sampling acceleration as an optimization problem. Shuai Wang et al. proposed Differentiable Solver Search Differentiable Solver Search for Fast Diffusion Sampling, which automatically learns the optimal timesteps and coefficients for a solver, outperforming hand-crafted mathematical methods.

Increasingly, Reinforcement Learning (RL) is used to align fast models with human preferences. Dengyang Jiang and team created DMDR Distribution Matching Distillation Meets Reinforcement Learning, combining distribution matching with RL to ensure that 4-step models don't just generate images quickly, but generate better images than their teachers. Similarly, Yihong Luo et al. introduced Reward-Instruct Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation, which uses direct reward maximization to bypass expensive distillation losses.

Domain-Specific Efficiency

These acceleration techniques are being rapidly adapted for specialized tasks:

Papers in this version

Showing 60 of 164 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_accelerating_diffusion_sampling_via_dist_38cf716d,
  title = {Accelerating Diffusion Sampling via Distillation and Trajectory Rectification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/accelerating-diffusion-sampling-via-distillation-and-trajectory-rectification/rdr_38cf716d70854f298656df3fceee3923},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_accelerating_diffusion_sampling_via_dist_38cf716d_v1,
  title = {Accelerating Diffusion Sampling via Distillation and Trajectory Rectification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/accelerating-diffusion-sampling-via-distillation-and-trajectory-rectification/rdr_38cf716d70854f298656df3fceee3923/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}