Understanding Visual Concepts Across Models

TL;DR

This paper analyzes how large multimodal models learn and represent new visual concepts through word embeddings, revealing their model-specific nature and non-transferability across different models and tasks.

Contribution

It provides a large-scale analysis of visual concept embeddings, demonstrating their non-transferability and the existence of perturbative solutions that can generate or classify arbitrary concepts.

Findings

Embeddings are model-specific and non-transferable.

Perturbations within an epsilon-ball can generate or classify arbitrary concepts.

Popular soft prompt-tuning methods find these perturbative solutions.

Abstract

Large multimodal models such as Stable Diffusion can generate, detect, and classify new visual concepts after fine-tuning just a single word embedding. Do models learn similar words for the same concepts (i.e. <orange-cat> = orange + cat)? We conduct a large-scale analysis on three state-of-the-art models in text-to-image generation, open-set object detection, and zero-shot classification, and find that new word embeddings are model-specific and non-transferable. Across 4,800 new embeddings trained for 40 diverse visual concepts on four standard datasets, we find perturbations within an $\epsilon$-ball to any prior embedding that generate, detect, and classify an arbitrary concept. When these new embeddings are spliced into new models, fine-tuning that targets the original model is lost. We show popular soft prompt-tuning approaches find these perturbative solutions when applied to visual concept learning tasks, and embeddings for visual concepts are not transferable. Code for reproducing our work is available at: https://visual-words.github.io.