Shortcut Diffusion Training With Cumulative Consistency Loss: An Optimal Control View

Area of Science:

• Artificial Intelligence
• Machine Learning
• Generative Models

Background:

• Iterative denoising methods (diffusion/flow) excel at generation but are inefficient.
• Current methods for faster generation, like shortcut models, compromise quality.
• Few-step generation requires larger simulation steps for efficiency.

Purpose of the Study:

• To improve the efficiency and quality of few-step generative models.
• To address the performance degradation seen in existing fast generation techniques.
• To develop a novel loss function for enhanced generative model training.

Main Methods:

• Formulating few-step generation as a controlled generative process.
• Generalizing self-consistency loss to a cumulative self-consistency loss (CSL).
• Connecting the approach to optimal control and reinforcement learning principles.

Main Results:

• The proposed CSL significantly improves one- and few-step generation quality.
• Achieved higher generation quality under the same training budget compared to prior methods.
• Demonstrated the effectiveness of cumulative penalization for trajectory alignment.

Conclusions:

• Cumulative self-consistency loss enables high-quality, efficient few-step generation.
• The optimal control perspective provides a theoretical foundation for improved generative models.
• This work opens new avenues for few-step generation using reinforcement learning concepts.