Time step self-adaptive diffusion model post-training quantification method

Through the time-step adaptive quantization method, the problem of the gap in the channel activation value range and iterative reasoning characteristics in the diffusion model are not considered, and more efficient quantization and performance improvement are achieved.

CN119722506AActive Publication Date: 2025-03-28BEIHANG UNIV
9 Cites 3 Cited by

Patent Information

Application Number
CN202411769042.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-12-04
Publication Date
2025-03-28
Estimated Expiration
2044-12-04

AI Technical Summary

Technical Problem

The existing quantization method ignores the gap in the activation value range between channels in the diffusion model and does not consider iterative reasoning characteristics, resulting in quantization error and performance degradation.

Method used

A time-step adaptive diffusion model post-training quantization method is proposed. Through the initialization of the fine-grained time domain and the quantitative reconstruction method to eliminate deviation errors, dynamic adaptive selection of quantizers is enhanced to enhance the model's perception and adaptability to channel activation values.

Benefits of technology

It effectively reduces quantization errors, improves the performance and robustness of the diffusion model under low quantization bit widths, and ensures that the model can perform fixed-point number inference on hardware.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention belongs to the field of diffusion model quantification, and provides a time step self-adaptive diffusion model post-training quantification method, which comprises the following steps: S1, obtaining a pre-trained full-precision diffusion model, and determining a sampler and a sampling strategy; s2, generating an image based on the full-precision diffusion model obtained in the step S1 and the sampler, and sampling to obtain a quantized check set after training; s3, establishing a diffusion model, and performing quantization parameter initialization on the diffusion model; s4, performing reconstruction training on the diffusion model; and S5, performing performance evaluation on the diffusion model subjected to reconstruction training. The method can effectively solve the problems that an existing quantization method neglects an inter-channel activation value range gap, does not consider the influence of iterative reasoning specificity on quantization reconstruction and performance collapse of a quantized model under a low quantization bit width, and gives full play to the advantages of a time step perceived post-training quantization method. And the quantized diffusion model has better performance and robustness.
Need to check novelty before this filing date? Find Prior Art