What question did this study set out to answer?

The aim is to distill pre-trained diffusion and flow matching models into efficient one-step generators while preserving image quality.

One-Step Diffusion and Flow Distillation through Implicit Generator Matching

View Full Paper

Ask AI

Bookmark

The aim is to distill pre-trained diffusion and flow matching models into efficient one-step generators while preserving image quality.
Developed a systematic approach called Implicit Generator Matching (IGM) for model distillation.
Utilized the Implicit Gradient Theorem to provide efficient optimization of student generator models.
Maintained data-free methodologies, eliminating the need for training images.
Achieved a FID score of 2.06 on CIFAR10 using the diffusion-based SIM model.
Set a flow-model record with a FID score of 3.08 using the flow-based FGM model.
Obtained a leading 6.42 aesthetic score in text-to-image models, surpassing SDXL-TURBO.
Achieved a competitive GenEval score of 0.65 against multi-step models like Hyper-SD3.

No takes yet. Share an insight, caveat, or question.

Demonstrates efficient distillation of generative models, achieving high image quality without training data.

Huang et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark