AI Paper Atlas

Research map / Vision-language and multimodal / Cross-modal learning

MAGIC: Multi-Granularity Language-Informed Image Clustering

Xiaohan Zhang, Chao Zhang, Chunlin Chen, Huaxiong Li

ICML 2026 · Vision-language and multimodal › Cross-modal learning · approach: Medical multimodal fusion

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 changedWhereMeasuredEffect
add multi-modal language information (MAGIC vs k-means on image CLIP) (ablation)MAGIC · Five-dataset averageAVG: 65.6 -> 87.2Improved
use VLM descriptions instead of WordNet nouns (ablation)MAGIC · Five-dataset averageAVG: 84.3 -> 87.2Improved
add balance loss L_BL (ablation)MAGIC · CIFAR-20ACC: 27.8 -> 66.1Improved
add cross-entropy alignment loss L_CE (ablation)MAGIC · ImageNet-DogsACC: 90.1 -> 93.1Improved
add inter-modal contrastive distillation to intra-modal (ablation)MAGIC · CIFAR-20ACC: 55.6 -> 66.1Improved
fuse fine and coarse text embeddings instead of fine only (ablation)MAGIC · CIFAR-20ACC: 61.0 -> 66.1Improved
add semantic adapters (image and text) (ablation)MAGIC · CIFAR-20ACC: 57.1 -> 66.1Improved
add residual connection to semantic adapters (ablation)MAGIC · CIFAR-20ACC: 66.1 -> 59.9Hurt
bottleneck ratio 1/4 vs 1 or 1/8 (ablation)MAGIC · CIFAR-20ACC: 62.9 -> 66.1Improved
use cross-granularity attention instead of addition or concatenation (ablation)MAGIC · ImageNet-Dogs, CIFAR-20clustering metricsImproved
concatenate image embedding into text cluster head (consensus representation) (ablation)MAGIC · ImageNet-Dogs, CIFAR-20clustering metricsImproved
use Qwen2.5-VL instead of BLIP-2 or WordNet as language generator (ablation)MAGIC · ImageNet-Dogs, CIFAR-20clustering metricsImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy