Text-to-image retrieval usually optimizes for relevance only. In practice you often care about quality too: more aesthetic photos, fewer blurry or low-IQA images, or a custom trade-off. We call this Quality-Controllable Retrieval (QCR), a new setting where retrieval can be explicitly conditioned on user-defined quality requirements.
We propose Quality-Conditioned Query Completion (QCQC), a query completion framework that leverages LLMs to enrich short queries with quality-aware descriptive details. Specify desired quality (e.g., aesthetic, relevance, image quality), and QCQC completes your query so retrieval returns results that match both meaning and quality.
Quality control — Describe desired quality as the condition; no separate filters or post-hoc ranking.
Multi-dimensional quality — Aesthetic, image quality (IQA), and relevance, composable in one framework (adapt to any quality definition).
Reproducible — MS-COCO workflow, clear data pipeline, and training/inference scripts.
Overview
We use MS-COCO and GPT-2 as the running example: download data, build a search index, generate auxiliary quality scores (aesthetic, IQA, relevance), tokenize the data, train the QCQC model, and then run retrieval. The steps below walk through the full pipeline.
Qualitative examples of quality-conditioned retrieval:
Results 1
Results 2
Quality-conditioned retrieval examples (1)
Quality-conditioned retrieval examples (2)
Citation
If you use this code or idea in your work, please cite:
bibtex
1@inproceedings{JianglinQCQC2026,
2 title = {Seeing Through Words: Controlling Visual Retrieval Quality with Language Models},
3 author = {Jianglin Lu and Simon Jenni and Kushal Kafle and Jing Shi and Handong Zhao and Yun Fu},
4 booktitle = {The Fourteenth International Conference on Learning Representations (ICLR)},
5 year = {2026},
6 url = {https://openreview.net/forum?id=yOEmEXmbV8},
7}
Acknowledgement
We use the following open-source projects and thank the authors:
