SMD: Microsoft Research Reduces Image and Video Generation to Four Diffusion Steps

Microsoft Research introduces SMD, a distillation method that accelerates image, video, audio, and 3D generation into four steps.

Accelerating image and video generation without compromising quality remains one of the main challenges facing diffusion models. These systems gradually construct outputs through repeated computational steps, and reducing those steps can produce less detailed images, inconsistent movement, or diminished diversity. Researchers affiliated with Microsoft Research, New York University, and the Hong Kong University of Science and Technology propose another approach with Simplified Matching Distillation (SMD). Presented in an October 2026 research paper, the method adapts existing models to generate content in as few as four steps without adding auxiliary networks at inference time. The team evaluates it across image generation and editing, video, audiovisual synthesis, and 3D, including models with up to 33 billion parameters. Fewer Computations Without Rebuilding the Generator

Diffusion and flow-matching models generally begin with noise and gradually transform it into an image, video, or another type of output. Each step updates the result based on another prediction from the model. The process may require dozens or hundreds of sequential network evaluations, increasing both latency and computational cost.

Distillation aims to reduce this burden by using a pretrained teacher model to train a student capable of producing comparable results with fewer operations. Some existing approaches, however, require additional neural networks, complex training objectives, or specialized generator architectures. Others can reduce the diversity of generated outputs by favoring a narrower range of compositions.

SMD simplifies the problem through on-policy velocity regression. Instead of requiring the student to reproduce every intermediate prediction along the teacher's sampling trajectory, it learns larger transitions from the teacher's local velocity predictions. Training takes place on trajectories generated by the student itself, allowing the model to learn from the states it will encounter during inference.

The researchers found that this relatively simple approach already performs well, but can become unstable when trajectories involve complex changes, especially in video. To address that problem, SMD introduces two additional mechanisms that operate only during training. Learning Which Directions to Follow and Avoid

The first mechanism, called Pivot, is a small auxiliary network that represents local variations within each coarse generation step. It allows training to account for changing teacher predictions without altering the final point predicted by the student. The additional network is discarded after training, so it does not increase inference requirements.

The second mechanism adds a repulsive training signal. Alongside the teacher prediction it should imitate, the student receives a weaker prediction from the same teacher as a negative reference. Training encourages it to move away from that less desirable direction while remaining close to the intended target. The objective is to improve visual quality and preserve output diversity without adversarial training.

Both mechanisms refine the student during distillation. Once training is complete, neither Pivot nor the teacher is needed for generation. The student retains its original architecture and performs its four sampling steps directly. SMD is therefore a training method for existing model families rather than a new universal image or video generator. FLUX.1 and Qwen-Image Deliver Strong Results in Four Steps

The researchers initially evaluate SMD using FLUX.1-dev, with 12 billion parameters, and Qwen-Image, with 20 billion. Both are distilled into four-evaluation image generators and compared with other few-step approaches.

On GenEval, which measures the alignment between prompts and generated compositions, SMD's FLUX.1-dev student achieves 0.715, compared with 0.675 for π-Flow and 0.664 for the original teacher using 50 evaluations. Its DPG-Bench score reaches 84.8, also exceeding the methods included in the comparison. Under these test conditions, the distilled model improves selected quality metrics while requiring substantially fewer evaluations.

With Qwen-Image, SMD reaches 0.873 on GenEval and 88.7 on DPG-Bench. The original teacher scores 0.874 and 88.3, respectively. The distilled model also improves text rendering on OneIG-Bench-EN, scoring 0.944 compared with the teacher's 0.889. Gains are not universal, however: its diversity score remains below that of the original model.

These results are specific to the study's evaluation settings. Reducing the number of model evaluations from fifty to four can significantly reduce computational work, but actual generation speed still depends on architecture, supporting components, and hardware. Image Editing and 3D Generation

The team also applies SMD to Qwen-Image-Edit-2511, which modifies images based on instructions. Evaluated operations include inserting, replacing, and removing objects, editing text, and changing visual appearance. At four evaluations, SMD scores 4.51 on ImgEdit-Bench, compared with 4.48 for Qwen-Image-Edit Lightning. Its overall GEdit-Bench-EN score improves from 8.14 to 8.19, although performance varies by editing operation.

The method also transfers to 3D generation. With Hunyuan3D-2.1, the researchers reduce the shape-generation stage to four evaluations. With TRELLIS.2, they separately distill its high-resolution geometry and texture stages to four evaluations each. The examples preserve recognizable object structures and surface details, whereas simply forcing the original teacher to run in four steps can produce fragmented geometry.

That distinction matters: for TRELLIS.2, the four-step count applies independently to two distilled stages, not to the entire asset-generation pipeline. Earlier stages remain unchanged. The 3D results are also presented primarily through visual comparisons rather than a comprehensive quantitative benchmark. Preserving Motion in Video Generation

V