Self-Guidance: Boosting Flow and Diffusion Generation on Their Own.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• High-quality image generation from text requires effective guidance strategies for diffusion and flow-based models.
• Existing guidance methods often necessitate specific training or rely on model architecture biases, limiting flexibility.

Purpose of the Study:

• To introduce a novel guidance strategy, Self-Guidance (SG), that enhances image generation quality without requiring model retraining.
• To develop an efficient variant, SG-prev, for improved computational performance.

Main Methods:

• Self-Guidance (SG) leverages the sampling score function at different noise levels to detect and suppress low-quality samples.
• SG-prev optimizes efficiency by reusing previous diffusion step outputs, avoiding extra forward passes.
• Experiments were conducted on text-to-image and text-to-video tasks using UNet and transformer architectures.

Main Results:

• SG significantly improves generation quality, outperforming existing algorithms on metrics like FID and Human Preference Score across various models (e.g., Stable Diffusion 3.5, FLUX).
• SG-prev achieves comparable results to SG with 50% greater efficiency.
• Both SG and SG-prev demonstrate a notable reduction in human body artifacts, improving the generation of hands, faces, and arms.

Conclusions:

• Self-Guidance offers a flexible, plug-and-play solution for enhancing diffusion and flow-based model generation quality.
• The efficient SG-prev variant provides a practical approach for real-time applications.
• SG methods show promise in mitigating common generation artifacts, particularly in complex human anatomy.