tianweiy/DMD2

Health v2 · maintenance only

Flags: no_releases no_license

Adoption not part of the score

1438 stars · 76 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

DMD2 is the official PyTorch implementation of Improved Distribution Matching Distillation, a NeurIPS 2024 method that distills diffusion models into fast one- or few-step image generators. It removes costly regression losses, adds a GAN loss, and supports distilling models like SDXL for high-quality few-step text-to-image synthesis.

Use cases

• distill a diffusion model into a one-step generator
• generate images in a few steps instead of many
• speed up SDXL text-to-image inference
• reproduce DMD2 research results on ImageNet and COCO
• train a fast image generation model with a GAN loss
• reduce inference cost of diffusion models by 500x

When to choose

• you need fast few-step text-to-image generation with near-teacher quality
• you want to distill SDXL or ImageNet diffusion models into efficient generators
• you are researching distribution matching distillation techniques

When to avoid

• you just want to run a pretrained model without training or distillation
• you lack GPU resources for training large diffusion models
• you need a production-ready inference server rather than research code

Facets

library · maturity active

machine-learning deep-learning llm-training image-processing machine-learning deep-learning image-processing artificial-intelligence python diffusion-models model-distillation text-to-image few-step-generation research-code sdxl gan-loss gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="tianweiy/DMD2")