Diffusion Image Editing With Dual Guidance Scale Branches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Denoising Diffusion Implicit Models (DDIM) inversion in image editing suffers from approximation errors due to asymmetric guidance scales in classifier-free guidance, leading to undesired performance in image generation and editing.
Innovation Solution
Implementing a symmetric guidance scale for the source branch and maintaining a larger scale for the target branch in the dual branch framework of DDIM inversion to reduce approximation errors, while keeping the framework simple.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If asymmetric guidance scales are used in DDIM inversion for image editing, then the framework maintains simplicity, but approximation errors accumulate resulting in undesired performance
Solution Approach 1:
The patent divides the guidance scale into two separate components: a first guidance scale for the source image branch and a second guidance scale for the target image branch. This segmentation allows each branch to operate with optimized guidance scales independent of each other, reducing approximation errors while maintaining framework simplicity.
Solution Approach 2:
The patent applies different guidance scale values to different branches of the inversion process. The source image branch uses a first guidance scale optimized for preserving source image characteristics, while the target image branch uses a second guidance scale optimized for achieving desired editing results. This local differentiation improves overall image editing fidelity.
2Manufacturing precision
If symmetric guidance scales are used in both source and target branches, then approximation errors are minimized, but the flexibility to control different editing aspects is reduced
Solution Approach 1:
The patent introduces dynamic control by allowing different guidance scale values in the source and target branches. This dynamic configuration enables flexible adjustment of each branch's behavior independently, providing adaptability for different editing scenarios while maintaining high fidelity through optimized guidance scales.
Solution Approach 2:
The patent changes the guidance scale parameter differently for the source image branch and target image branch. By allowing independent parameter optimization for each branch, the system achieves both high editing fidelity and flexible control over different aspects of the image editing process.
3Manufacturing precision
If larger guidance scale is used in target branch, then target image generation quality is improved, but source image recovery accuracy may be compromised
Solution Approach 1:
The patent segments the guidance scale application into two distinct parts: a first guidance scale for source image recovery and a second guidance scale for target image generation. This segmentation allows the target branch to use larger guidance scales for improved generation quality without compromising source image recovery accuracy, as each branch operates with its own optimized scale.
Solution Approach 2:
The patent applies locally optimized guidance scales to different branches. The source image branch uses a guidance scale optimized for accurate recovery, while the target image branch uses a larger guidance scale optimized for high-quality generation. This local optimization resolves the trade-off between recovery accuracy and generation quality.
Data Source
AI summary
The present disclosure provides a method for image generation, an electronic device and a non-transitory storage medium. The method includes the following steps: a source image, a source prompt, a target prompt, a first weight and a second weight are obtained, the source prompt is used for providing a source text condition corresponding to the source image, the target prompt is used for providing a target text condition to obtain a target image, the first weight is a weight of a guidance scale shared for inversion and generation in a source branch, the second weight is a weight of a guidance scale kept for generation in a target branch; inversion for image editing is performed to generate the target image based on the source image, the source prompt, the target prompt, the first weight and the second weight.