Autoregressive text generation acceleration method based on parallel jacobian decoding

By employing a row-level activation strategy and a probabilistic acceptance mechanism in parallel Jacobi decoding, the inference efficiency bottleneck of autoregressive image generation methods in high-resolution scenarios is resolved, achieving efficient image generation while maintaining the stability of generation quality and ease of engineering implementation.

CN122199719APending Publication Date: 2026-06-12WESTLAKE UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
WESTLAKE UNIV
Filing Date
2026-02-12
Publication Date
2026-06-12

AI Technical Summary

⚠Technical Problem

Existing autoregressive image generation methods suffer from bottlenecks in inference efficiency, particularly in their inability to fully utilize the two-dimensional spatial structure and local correlations of images, leading to a decrease in iterative convergence speed and difficulty in effectively scaling up in high-resolution scenarios.

⚗Method used

A parallel Jacobi decoding-based approach is adopted, which processes multiple image tags in parallel in a single forward computation through a row-level activation strategy and a probabilistic acceptance mechanism. It leverages the two-dimensional spatial characteristics of images to dynamically expand and update draft tags in parallel, and uses row-causal attention masks for conditional probability verification.

🎯Benefits of technology

It significantly improves generation efficiency, reduces the overall decoding rounds, maintains stable generation quality, reduces training and deployment complexity, and has good versatility and scalability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122199719A_ABST
    Figure CN122199719A_ABST
Patent Text Reader

Abstract

The application discloses an autoregressive text-to-image acceleration method based on parallel Jacobian decoding, and belongs to the technical field of image generation. The method comprises inputting and initializing, dynamically constructing a draft label, parallel autoregressive prediction, and probabilistic draft label verification. In view of the problem that the existing Jacobian decoding mainly adopts one-dimensional draft sequence expansion, and the late-stage label converges slowly in the text-to-image task, the application utilizes the two-dimensional space correlation and attention locality of image labels, dynamically expands the draft in a two-dimensional label grid, and performs parallel iterative prediction and probability verification on the draft label, so that multiple labels are confirmed in parallel. Compared with the label-by-label serial decoding and one-dimensional expansion Jacobian scheme, the application can confirm more image labels in a single forward process, reduce the number of decoding iterations, reduce the inference delay, and keep the generated image quality stable while accelerating, so as to improve the overall generation efficiency and engineering practicability of the autoregressive text-to-image model.
Need to check novelty before this filing date? Find Prior Art