Research map / Label-efficient and robust learning / Recognition and long-tail
IRGen: Generative Modeling for Image Retrieval
Yidan Zhang, Ting Zhang, Dong Chen, Yujing Wang, Qi Chen, Xing Xie, Hao Sun, Weiwei Deng, Qi Zhang, Fan Yang, Mao Yang, Qingmin Liao and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use semantic RQ identifiers instead of random codes (ablation) | IRGen · In-shop Clothes | Precision@10: 75.4 -> 89.7 | Improved |
| use proposed semantic tokenizer instead of hierarchical k-means identifiers (ablation) | IRGen · In-shop Clothes | Precision@10: 81.7 -> 89.7 | Improved |
| use PQ codes as identifiers instead of residual quantization (ablation) | IRGen · In-shop Clothes | Precision@10: 89.7 -> 66.1 | Hurt |
| use RQ-VAE tokenizer with long sequences (ablation) | IRGen · In-shop Clothes | Precision | Hurt |
| increase identifier length from 2 to 4 (ablation) | IRGen · In-shop Clothes | Precision@1: 72.1 -> 92.4 | Improved |
| increase identifier length from 4 to 8 (ablation) | IRGen · In-shop Clothes | Precision@1: 92.4 -> 92.9 | No clear effect |
| add autoregressive decoder on top of RQ prefix-tree beam search (ablation) | IRGen · ImageNet | MAP@100: 56.7 -> 76.0 | Improved |
| add partial reconstruction losses to classification loss in tokenizer training (ablation) | IRGen tokenizer · In-shop Clothes | Precision@10: 85.1 -> 87.0 | Improved |
| IRGen generative retrieval versus FT-CLIP with ANN or linear scan | IRGen · In-shop Clothes | Precision@10: 66.8 -> 87.0 | Improved |
| IRGen versus ScaNN/SPANN baselines on recall at larger K | IRGen · In-shop Clothes | Recall@10: 97.8 -> 96.8 | Hurt |
| retrieve with IRGen on newly added unseen data without retraining | IRGen · ImageNet unseen 5% data | Precision@1: 70.6 -> 77.0 | Improved |
Closest papers by meaning
- Image Generation Diversity Issues and How to Tame Them (CVPR 2025 · similarity 0.87)
- Generative Retrieval Meets Multi-Graded Relevance (NeurIPS 2024 · similarity 0.87)
- GenIR: Generative Visual Feedback for Mental Image Retrieval (NeurIPS 2025 · similarity 0.87)
- Learning to Rank in Generative Retrieval (AAAI 2024 · similarity 0.87)
- PQPP: A Joint Benchmark for Text-to-Image Prompt and Query Performance Prediction (CVPR 2025 · similarity 0.86)
- TIGeR: Unifying Text-to-Image Generation and Retrieval with Large Multimodal Models (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
