Progressive classifier-free guidance for image synthesis

By employing progressive classifier-free guidance and weakened condition embeddings, the method addresses the issue of token domination in personalized text-to-image synthesis, resulting in more diverse and contextually accurate image generation.

WO2025019263A1PCT designated stage expired Publication Date: 2025-01-23GOOGLE LLC
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
PCT/US2024/037573
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-07-14
Filing Date
2024-07-11
Publication Date
2025-01-23

AI Technical Summary

Technical Problem

Existing personalized text-to-image synthesis networks are prone to overfitting the target subject from the input image, leading to generated images that ignore important context in the input text.

Method used

The use of progressive classifier-free guidance, which involves generating a weakened condition embedding by replacing the token representing the target subject with a generic word, and performing multiple rounds of classifier-free guidance to balance context consistency and subject fidelity.

Benefits of technology

This approach effectively reduces token domination, allowing for more diverse and contextually accurate image synthesis, where the generated images include unique features of the target subject without losing the context from the text condition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2024037573_23012025_PF_FP_ABST
    Figure US2024037573_23012025_PF_FP_ABST
Patent Text Reader

Abstract

Methods and systems for generating an image are disclosed herein. The method can include receiving an input condition embedding from an encoder, the input condition embedding including a text condition and a token representing a target subject for a generated image, and generating a weakened condition embedding by replacing the token representing the target subject with an embedding representing a generic word describing the target subject. The method also includes performing at least one first denoising step based on the weakened condition embedding to receive a first output, performing at least one second denoising step based on the input condition embedding and the first output to receive a second output, and generating a generated image based on at least the second output.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Yoga Dragon. A soft tube designed to hook and / or fix onto a yoga mat

    US62635136P0