Show Me Examples: Inferring Visual Concepts from Image Sets
| Source: Apple ML Research
Tags: Apple, VICIS, vision-language models, VLM, few-shot learning, image generation, ECCV, compositional generation
Apple ML Research reveals that state-of-the-art VLMs fail at inferring shared visual concepts from example image sets — they ignore visual context or default to biased outputs. The VICIS paper (ECCV 2026) introduces both the benchmark task and a training framework that learns concept-specific embeddings from image sets.
Details
Apple ML Research introduces VICIS (Visual Concept Inference from Sets), a new evaluation task for vision-language models — and a stark finding: current SOTA models fail at it. Given a small set of images sharing a concept (no text label) plus a query image, models should generate new images that preserve the concept while matching the query. State-of-the-art VLMs instead ignore the visual context or produce biased generations. The task is deliberately harder than few-shot prompting with text descriptions. The concept is communicated purely through example images, requiring the model to extract visual patterns without linguistic scaffolding — closer to how humans learn styles or aesthetics by example rather than definition. The Apple team proposes a training framework and architecture that learns concept-specific embeddings from image sets. Experiments on synthetic data and large-scale ImageNet/WordNet data show better accuracy and diversity, with generalization to unseen concepts and to sketch-mode inputs the model was not explicitly trained on. The practical stakes are in personalization: systems that could learn what a brand's visual identity or a user's style looks like from examples rather than text prompts. The paper appears at ECCV July 2026; lead authors are from Apple and LMU Munich.