One-Step Diffusion and Flow Distillation through Implicit Generator Matching.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Diffusion and flow matching models excel at generative tasks but require numerous sampling steps for high-quality image generation.
• The need for efficient generative models has driven research into distilling pre-trained models into faster, more streamlined versions.
• Existing distillation methods often struggle with the implicit field definitions of student generators compared to the explicit fields of teacher models.

Purpose of the Study:

• To introduce Implicit Generator Matching (IGM), a novel approach for distilling pre-trained diffusion and flow matching models into one-step generator models.
• To develop a data-free distillation method that preserves the generative capabilities of the original models without requiring training images.
• To address the challenge of distilling models with differing field definitions (explicit vs. implicit) using an effective gradient optimization technique.

Main Methods:

• Developed Implicit Generator Matching (IGM), a systematic approach for model distillation.
• Introduced the Implicit Gradient Theorem to provide an exact and efficient gradient for optimizing student generators.
• Aligned the implicit field of the student generator with the explicit field of the teacher model.

Main Results:

• Achieved state-of-the-art performance in one-step generator models across various benchmarks.
• Diffusion-based SIM model attained an FID score of 2.06 on CIFAR10.
• Flow-based FGM model set a new record with an FID score of 3.08 on CIFAR10.
• Distilled text-to-image models achieved leading scores: SIM distillation of PixArt-$\alpha$ yielded an aesthetic score of 6.42, and FGM distillation of SD3 achieved a 0.65 GenEval score.

Conclusions:

• IGM offers an effective and data-free method for distilling diffusion and flow matching models into highly efficient one-step generators.
• The Implicit Gradient Theorem is a key breakthrough enabling direct optimization of student generators with implicit fields.
• IGM-based models demonstrate superior performance and efficiency, setting new standards in generative modeling for both image and text-to-image synthesis.