Image Selection Using Language Model Attractiveness Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image selection algorithms in platforms like e-commerce and news aggregation fail to fully account for the rich dimensions of image quality, leading to under-discovery and under-utilization of long-tail and new products, as they primarily rely on engagement metrics such as clicks and views, which are inefficient for evaluating the appeal of images with models or lifestyle features.
Innovation Solution
A method that involves obtaining candidate images, generating a prompt for a language model, obtaining attractiveness ranks, and determining probability distributions to select a target image for display, allowing the model to comprehensively evaluate image features like human models and scenes, thereby improving the accuracy of ranking and giving more display opportunities to the most attractive images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If engagement metrics such as clicks and views are used to evaluate image appeal, then the selection process is simple and fast, but the evaluation accuracy is insufficient and fails to account for rich dimensions of image quality
Solution Approach 1:
The patent introduces an intermediary evaluation model that bridges the gap between simple engagement metrics and comprehensive image quality assessment. This model incorporates multiple dimensions including model attractiveness, scene quality, and composition factors, transforming raw engagement data into refined quality scores without requiring direct complex analysis of each image dimension.
Solution Approach 2:
The system dynamically adjusts evaluation parameters based on different contexts and image types. By modifying the weightings and thresholds of various quality dimensions according to the specific evaluation scenario, the system achieves high measurement precision while adapting to different complexity requirements without a fixed rigid structure.
2Productivity
If traditional algorithms rely on engagement metrics like clicks and views, then the implementation is straightforward, but long-tail and new products are under-discovered and under-utilized
Solution Approach 1:
The system performs preliminary evaluation of images using multiple quality dimensions before they accumulate sufficient engagement metrics. By pre-assessing model attractiveness, scene quality, and composition, the system identifies promising long-tail and new products early in their lifecycle, preventing information loss about potentially valuable content that hasn't yet gained traction.
3Measurement precision
If the system selects images based on limited metrics, then the selection process is fast, but the accuracy of ranking is insufficient
Solution Approach 1:
The evaluation process is segmented into distinct modular components: model attractiveness assessment, scene quality evaluation, composition analysis, and engagement metric integration. Each segment operates independently and can be processed in parallel, maintaining fast selection speed while achieving comprehensive and accurate ranking through the aggregation of multiple specialized evaluations.
Data Source
AI summary
Implementations of the present disclosure provide a method, device, and medium for determining an image for display. The method comprises obtaining a plurality of candidate images associated with an object. The method further comprises generating a prompt for a language model based on the plurality of candidate images. The method further comprises obtaining a plurality of attractiveness ranks corresponding to the plurality of candidate images by feeding the prompt to the language model. The method further comprises determining a plurality of probability distributions corresponding to the plurality of candidate images based on the plurality of attractiveness ranks. In addition, the method further comprises determining a target image for display from the plurality of candidate images based on the plurality of probability distributions.