Alibaba's CDM slashes diffusion steps to 4 without losing detail

Waiting for a diffusion model to generate an image can feel like watching paint dry. Standard pipelines grind through 20 to 50 denoising steps, easily burning 15 to 20 seconds per prompt. While distillation techniques like DMD brought those numbers down to 4 steps, they often left generations looking waxy, flat, and aggressively over-smoothed.

Alibaba introduced Continuous-Time Distribution Matching, or CDM, an acceleration method designed to bridge that gap. The core payoff is simple: it slashes the required sampling steps of diffusion models from 20 to 50 steps down to just 4, delivering roughly a 5x speedup while preserving image quality.

Ditching the plastic look without GAN bloat

Until now, few-step distillation largely hinged on Distribution Matching Distillation (DMD) and consistency models. Vanilla DMD operates on a few predefined discrete timesteps. Because of the mode-seeking nature of reverse KL divergence, this discrete setup routinely causes visual artifacts and plastic skin tones. To patch that loss of fidelity, researchers usually had to glue on complex external modules, such as GAN discriminators or reward models, making the entire setup cumbersome.

CDM attacks the root of the problem by migrating the framework from rigid discrete anchors to continuous optimization. Instead of monitoring a handful of fixed checkpoints, CDM uses a dynamic continuous schedule of random lengths, enforcing distribution matching at arbitrary points along the sampling path.

On top of that, Alibaba introduced a continuous-time alignment objective that performs active off-trajectory matching. In plain English, the student model learns to extrapolate latents via its own velocity field, keeping delicate high-frequency textures intact without relying on heavy auxiliary reward networks.

Tested on SD3 and LongCat

The team tested CDM across multiple architectures, notably Stable Diffusion 3 (SD3-Medium) and LongCat-Image. Across aesthetic benchmarks and quality evaluations, CDM consistently beat out DMD2, eliminating the greasy plastic aesthetic while keeping fine details remarkably sharp.

For anyone with a decent consumer GPU, this shrinks generation time down to a few seconds, turning what used to be a coffee break into near real-time rendering. The paper and code are already available, with pre-configured support for SD3 and LongCat.

Why it matters

For creators and local developers, inference latency is the ultimate friction point. Squeezing diffusion down to 4 steps without butchering texture quality means rapid prototyping and interactive workflows can finally run locally on affordable hardware, rather than demanding dedicated cloud clusters.

From an engineering perspective, we think Alibaba's approach is remarkably clean. Rather than slapping on extra GAN patches or reward models to brute-force visual fidelity, CDM fixes the core sampling schedule itself. It proves that few-step distillation can be both lightweight and visually competitive.

The catch is ecosystem reach. While support for SD3-Medium and LongCat is a great start, the open-source community is eager to see whether CDM can scale smoothly to heavyweight architectures like Flux or Qwen, and whether complex prompt-following degrades when pushed to the absolute 4-step limit.