SMD

Simplifying and Improving On-Policy Distillation for Few-Step Diffusion

SMD is a simple and scalable method for few-step diffusion distillation across image, video, audio-video and 3D generation.

Paper · Project page · Code · Models · Datasets

This page is the project overview. Follow the links below for the released weights, inference instructions, workflows, datasets and original examples.

Explore the model family

The groups below organize models by task. Each student retains its own architecture and base-model requirements; SMD is the shared distillation method.

Task	Released models	Sampling steps
Text to image	FLUX.1-dev · Qwen-Image · SD3.5-Medium · FLUX.2 Klein 4B · Lens	4 for the students
Image editing	Qwen-Image-Edit-2511	4
Text to video	Wan2.1 1.3B · Wan2.1 14B · Causal Wan2.1 1.3B	4
Text to audio-video	MiniMax H3 T2VA: 4-step · 8-step	4 / 8
Reference to audio-video	MiniMax H3 Ref2VA: 4-step · 8-step	4 / 8
Image to 3D	Hunyuan3D-2.1 Shape · TRELLIS.2 Shape and Texture	4 per distilled stage

SD3.5 includes six variants: self, clipscore, geneval, ocr, pickscore and multi. FLUX.2 Klein 4B includes teacher-4b and teacher-9b; Lens includes teacher-lens and teacher-flux2-9b, plus the original Lens base. TRELLIS.2 keeps separate shape and texture adapters in one repository.

H3 in ComfyUI

Native ComfyUI releases provide setup instructions and ready-to-use workflows:

Task	4 steps	8 steps
T2VA	ComfyUI package	ComfyUI package
Ref2VA	ComfyUI package	ComfyUI package

Datasets and resources

Repository	Contents
H3 T2VA 100K	100,000 cached text-to-audio-video conditions
H3 Ref2VA 20K	20,000 reference-to-audio-video training instructions and their cached conditions
Categorized Prompts 150K	Categorized training prompts
Prompt Sample Resources	Fixed sampling inputs, cached conditions and preview resources

Get started

1. Choose a model for your task from the table above.
2. Open its repository and follow the standalone inference or ComfyUI instructions.
3. Use its provided example case and sampling settings, then adapt the inputs for your own task.

LoRA releases require the matching base model. Follow each package's pinned base revision and sampling settings. For training and export, use the SMD code repository.

Examples

Image generation	Audio-video generation	Image to 3D

Each model repository includes worked examples with original outputs and generation settings. Explore the image and video examples, Hunyuan3D mesh example, and TRELLIS.2 geometry and textured turntables.

Licenses and attribution

Models and datasets retain the licenses and attribution documented in their respective repositories. See those model and dataset cards for the terms that apply to each download.