✦
Summary
CDM, or Continuous-Time Distribution Matching, is a few-step diffusion distillation method for high-fidelity image generation. It is designed to make diffusion models generate strong images with very few inference steps, reducing latency while preserving texture, detail, and prompt alignment. The project focuses on distilling large image generators into faster samplers without relying on extra GAN losses or reward-model objectives.
Key Features
Distills diffusion models for high-quality few-step image generation.
Uses continuous-time distribution matching instead of only discrete schedule matching.
Avoids auxiliary GAN or reward-model objectives in the core distillation setup.
Improves fidelity at low inference-step counts such as four NFE.
Targets modern text-to-image diffusion backbones including SD3-style and LongCat-style models.
Helps reduce image-generation latency and compute cost.
Preserves sharper textures and fine-grained visual details in few-step outputs.
Supports research into efficient diffusion deployment and distillation.
About CDM
The method migrates distribution matching distillation from discrete schedules into a continuous-time optimization framework. This lets CDM better align the student generator with the teacher distribution across the diffusion trajectory, improving few-step quality at low numbers of function evaluations. The approach is demonstrated on modern image-generation backbones such as SD3-Medium and LongCat-style image models, where four-step generation can retain sharper details than competing distillation baselines.
For developers and researchers, CDM is useful when image quality and speed both matter. It can inform faster text-to-image products, local generation pipelines, and research on efficient diffusion sampling. The product is best understood as a model-distillation technique that improves the deployment economics of diffusion systems by making high-quality generation possible with fewer denoising steps.