NiCI-Pruning: Enhancing Diffusion Model Pruning via Noise in Clean Image Guidance.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Diffusion probabilistic models achieve high-quality image generation but are computationally expensive.
• Existing model compression techniques like pruning are less effective for diffusion models due to their iterative nature.
• Resource-limited deployment of diffusion models necessitates efficient compression strategies.

Purpose of the Study:

• To develop an effective pruning method for compressing diffusion models.
• To address the challenge of applying pruning to iterative diffusion processes.
• To enable the use of diffusion models in resource-constrained environments.

Main Methods:

• Proposing NiCI-Pruning (Noise in Clean Image Pruning), a method that uses noise predicted from clean images for reconstruction loss.
• Employing Taylor expansion for effective parameter importance evaluation within the reconstruction loss.
• Introducing an interval sampling strategy with a timestep-weighted schema to mitigate noise from later timesteps.

Main Results:

• NiCI-Pruning demonstrates superior performance in compressing diffusion models.
• The method achieves an average reduction of 30.4% in FID score increase across five datasets compared to state-of-the-art methods.
• Experimental results validate the effectiveness and superiority of the proposed NiCI-Pruning approach.

Conclusions:

• NiCI-Pruning offers an effective solution for compressing diffusion models, making them suitable for resource-limited scenarios.
• The novel approach successfully adapts pruning techniques to the iterative nature of diffusion models.
• The availability of code and models facilitates further research and application.