What Does DisParQ Change for Vision Models That Learn Without Labels?

A new method called DisParQ learns to break images into discrete, human-inspectable parts without using any class labels or language supervision. It builds on a frozen DINOv2 backbone and matches its 83.2% top-1 accuracy on ImageNet linear probing while enabling cross-category part retrieval.

How each patch gets exactly one concept

DisParQ stands for Discrete Parts with Quantized attributes. The method takes a powerful vision-only self-supervised backbone — DINOv2 — and freezes it. No class labels. No language supervision. A learnable prototype dictionary stores concepts. Every image patch is assigned to exactly one concept from that dictionary. Only a sparse subset of concepts activates for any given image. This sparsity keeps the representation readable. The dictionary itself is learned, not fixed. The backbone stays frozen. The concepts emerge from the backbone's own representation. The paper says this assignment is discrete and spatially grounded.

Where quantized attributes come from

A single concept like "wheel" varies across images. Some wheels are spoked. Some are solid. Some are large. DisParQ learns continuous residuals alongside each concept to capture that variation. Then it quantizes those residuals into discrete attributes. The attributes are not free-form text. They are discrete codes. A spatial decoder reconstructs the backbone's representation using only the concepts and their quantized attributes. Successful reconstruction means the discrete representation preserves the backbone's information. The paper reports that this reconstruction works.

What seven benchmarks show

The authors evaluate on seven datasets. General recognition: ImageNet, PartImageNet, Places. Fine-grained: CUB, Cars, Dogs, Flowers. On ImageNet linear probing DisParQ reaches 83.2% top-1. That matches the frozen DINOv2 teacher. Concept consistency beats language-aligned models. Fine-grained recognition stays competitive. Cross-category part-based retrieval becomes possible.

Missing details

The paper is under review. Submission date: 7 Oct 2026. File size: 16,701 KB. Seven authors: Adam Pardyl, Siddhartha Gairola, Sukrut Rao, Adam Wróbel, Bartosz Zieliński, Bernt Schiele, Dawid Rymarczyk. No downstream fine-tuning results. No comparison with fully supervised part detectors. No compute budget or training time. No code release mentioned. The method assumes a frozen DINOv2 backbone; other backbones are not tested. Prototype dictionary size is not specified in the abstract. Sparsity level is not quantified there either. Quantization scheme for attributes is not detailed in the summary.

One sentence

The method learns parts without labels.