Adaptive Guidance Scaling for Text-Guided Image Editing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning models for text-guided image editing rely on manually defined guidance scales, which are subjective and lack an objective method to determine optimal scales, leading to inconsistent generation quality.

Innovation Solution

Implement a guidance machine learning model that dynamically generates guidance scales based on input images and prompts, using latent tensors and embeddings to iteratively refine the editing process.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If manually defined guidance scales are used, then the model can perform text-guided image editing, but the generation quality is inconsistent due to subjective scale selection

Engineering Contradiction:

Improvegeneration qualityVSAvoidsubjective scale selection

Core Design Contradiction:

Manufacturing precisionVSEase of operation

Solution Approach 1:

The system uses the diffusion model itself to generate guidance scales through reinforcement learning, eliminating the need for external manual definition. The model learns to predict optimal guidance scales automatically based on the input image and text prompt, making the system self-sufficient in determining appropriate scale values for different editing scenarios.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The guidance scale parameter is transformed from a fixed manual value into a dynamic variable that changes based on the specific input conditions. The reinforcement learning model learns to adjust the guidance scale parameter adaptively according to the image content and text prompt requirements, optimizing the balance between prompt adherence and image quality for each specific case.

Inventive Principle:

Principle #35Parameter changes

2Adaptability or versatility

If fixed guidance scales are used, then the model structure remains simple, but the model cannot adapt to different editing scenarios

Engineering Contradiction:

Improveadaptability to different scenariosVSAvoidmodel structure

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

A reinforcement learning model is introduced as an intermediary component that bridges the gap between the diffusion model and the guidance scale parameter. This intermediary learns the optimal mapping from input conditions to appropriate guidance scales, enabling the system to adapt to different scenarios without fundamentally changing the core diffusion model architecture.

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The guidance scale transitions from a static fixed value to a dynamic parameter that varies with each input. The reinforcement learning model continuously adjusts the guidance scale based on the specific editing scenario, allowing the system to adapt its behavior to different images and text prompts while maintaining a relatively simple overall structure.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS12579713B2Text-guided image editing by learning guidance scales via reinforcement learning

Publication Date: 2026.03.17 QUALCOMM INC

AI summary

Certain aspects of the present disclosure provide techniques and apparatus for improved machine learning. In an example method, a first latent tensor generated during a first iteration of processing data using a denoising backbone of a diffusion machine learning model is accessed. A guidance scale is generated based on processing the first latent tensor using a guidance machine learning model. A second latent tensor is generated during a second iteration of processing data using the denoising backbone based on the first latent tensor and the first guidance scale, and an output from the diffusion machine learning model is generated based at least in part on the second latent tensor.