Research map / Label-efficient and robust learning / Recognition and long-tail

Learning to Align Generative Appearance Priors for Fine-grained Image Retrieval

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GAPan, which uses a RealNVP-style normalizing flow to model class-conditional Gaussian appearance priors over retrieval features and samples inverse-flow anchors to align embeddings for fine-grained image retrieval on unseen categories.

Problem

Fine-grained image retrieval models trained with category labels overfit to seen-category semantics and generalize poorly to unseen subcategories.

Approach

Reinterpret category labels as distribution-level appearance priors: a normalizing flow maps features to a latent space with class-conditional Gaussians (exact likelihood), and inverse-flow samples from high-density regions serve as appearance-aware anchors for a contrastive alignment loss.

Key results

• GAPan reaches Recall@1 of 89.1% on CUB-200-2011, 1.9 points above LaFG (87.2%).
• On Stanford Cars 196, GAPan reaches Recall@1 of 92.6%, 1.1 points above LaFG (91.5%).
• On Stanford Online Products, GAPan reaches Recall@1 of 88.4%, above LaFG (+1.3) and DPHM (+3.6).

Surprising findings

• Replacing the flow with a standard encoder-decoder MLP lowers CUB Recall@1 from 89.1% to 84.9%, which the authors read as evidence for invertibility.

Limitations

• Performance may inherit biases when training data are long-tailed or lack visual diversity.
• Evaluated only on three benchmarks with a single backbone, ViT-B/16.

Experimental setup

Datasets: CUB-200-2011 Stanford Cars 196 Stanford Online Products ImageNet-21K

Benchmarks: CUB-200-2011 Stanford Cars 196 Stanford Online Products

Metrics: Recall@1 Recall@2 Recall@4 Recall@8 Recall@10 Recall@100 Recall@1000

Baselines: LaFG DDML SEE DPHM DVA HypViT HIER NIR PNCA++

Builds on: RealNVP Glow NIA DDML

Closest papers by meaning

• FashionERN: Enhance-and-Refine Network for Composed Fashion Image Retrieval (AAAI 2024 · similarity 0.85)
• NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval (ICLR 2025 · similarity 0.85)
• Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms (NeurIPS 2024 · similarity 0.85)
• Composed Image Retrieval with Text Feedback via Multi-grained Uncertainty Regularization (ICLR 2024 · similarity 0.84)
• Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.84)
• FLAIR: VLM with Fine-grained Language-informed Image Representations (CVPR 2025 · similarity 0.84)