Embedding an input image to a diffusion model

Fine-tuning a diffusion model on a single target image with text prompts ensures consistent image variations that retain the target's identity, addressing the diversity issue in diffusion models and enhancing output quality.

AU2023226758B2Pending Publication Date: 2026-07-09ADOBE INC
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
AU · AU
Patent Type
Applications
Current Assignee / Owner
ADOBE INC
Filing Date
2023-09-08
Publication Date
2026-07-09

AI Technical Summary

Technical Problem

Diffusion models generate diverse outputs despite detailed descriptions due to stochastic noise, failing to consistently produce images resembling a target image.

Method used

Fine-tuning a diffusion model on a single target image to embed the input image in a latent text embedding space and generate variations that retain the target image's identity, incorporating text prompts for specific edits.

Benefits of technology

Produces consistent image variations that maintain the target image's characteristics while allowing for text-based modifications, reducing training time and enhancing output quality with multiple options.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000002_0000
    Figure 00000002_0000
  • Figure 00000038_0000
    Figure 00000038_0000
  • Figure 00000039_0000
    Figure 00000039_0000
Patent Text Reader

Abstract

2023226758 08 2023 Sep 1 / FIG. 1 110 115 100 105 120 Input Image Output Image A! !with! a!mustache" Input Prompt + 20 23 22 67 58 0 8 Se p 20 23 1 / 13 110 120 Output Image 115 "A person with many + a mustache" Input Prompt Input Image 105 100 FIG. 1 1 / 13 2023226758 08 120 115 + 100
Need to check novelty before this filing date? Find Prior Art