Continuous-Time Distribution Matching for Few-Step Diffusion Distillation
:::info Stub — Full Engineering Breakdown Coming This paper has a linked code implementation and was featured on Hugging Face Papers with 24 upvotes. A full breakdown with production viability rating, implementation notes, and honest limitations is being written. Subscribe to AI Letters → :::
| Authors | Tao Liu et al. |
| Year | 2026 |
| HF Upvotes | 24 |
| arXiv | 2605.06376 |
| Download | |
| Code | https://github.com/byliutao/cdm |
Abstract
Step distillation has become a leading technique for accelerating diffusion models, among which Distribution Matching Distillation (DMD) and Consistency Distillation are two representative paradigms. While consistency methods enforce self-consistency along the full PF-ODE trajectory to steer it toward the clean data manifold, vanilla DMD relies on sparse supervision at a few predefined discrete timesteps. This restricted discrete-time formulation and mode-seeking nature of the reverse KL divergence tends to exhibit visual artifacts and over-smoothed outputs, often necessitating complex auxiliary modules -- such as GANs or reward models -- to restore visual fidelity. In this work, we introduce Continuous-Time Distribution Matching (CDM), migrating the DMD framework from discrete anchoring to continuous optimization for the first time. CDM achieves this through two continuous-time designs. First, we replace the fixed discrete schedule with a dynamic continuous schedule of random length, so that distribution matching is enforced at arbitrary points along sampling trajectories rather than only at a few fixed anchors. Second, we propose a continuous-time alignment objective that performs active off-trajectory matching on latents extrapolated via the student's velocity field, improving generalization and preserving fine visual details. Extensive experiments on different architectures, including SD3-Medium and Longcat-Image, demonstrate that CDM provides highly competitive visual fidelity for few-step image generation without relying on complex auxiliary objectives. Code is available at https://github.com/byliutao/cdm.
Engineering Breakdown
Plain English
This paper proposes Continuous-Time Distribution Matching (CDM), a new approach for distilling diffusion models into fast, few-step versions. Instead of the existing discrete-timestep supervision used in Distribution Matching Distillation, CDM operates across the entire continuous trajectory, similar to how consistency methods work, but without requiring auxiliary networks like GANs to fix visual artifacts.
Key Engineering Insight
The core insight is that enforcing distribution matching continuously along the entire diffusion trajectory (rather than at sparse discrete timesteps) eliminates the mode-seeking artifacts and over-smoothing that plague current DMD approaches, removing the need for expensive auxiliary loss functions or reward models to fix output quality.
Why It Matters for Engineers
For engineers shipping diffusion-based products, this directly impacts inference speed and model size: fewer timesteps mean faster generation, but current distillation methods require kludges like GAN losses to maintain visual quality. CDM potentially gives you both speed and quality without these hacks, reducing complexity in your distillation pipeline and making faster models viable for production.
Research Context
Diffusion model acceleration has two main tracks: consistency methods (which enforce self-consistency but are complex) and DMD (which is simpler but produces artifacts). This paper bridges them by taking DMD's matching philosophy and extending it continuously, creating a middle ground that should combine the simplicity of DMD with the output quality of consistency approaches, advancing the practical viability of ultra-fast diffusion inference.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::