CogVLM: Visual Expert for Pretrained Language Models

TL;DR

CogVLM introduces a trainable visual expert module that enables deep vision-language feature fusion in pretrained language models, achieving state-of-the-art results across multiple cross-modal benchmarks without performance loss on NLP tasks.

Contribution

The paper proposes CogVLM, a novel visual language foundation model with a trainable visual expert module that enhances deep fusion of vision and language features in pretrained models.

Findings

Achieves state-of-the-art on 10 cross-modal benchmarks.

Ranks 2nd on VQAv2, OKVQA, TextVQA, and COCO captioning.

Surpasses or matches larger models like PaLI-X.

Abstract

We introduce CogVLM, a powerful open-source visual language foundation model. Different from the popular shallow alignment method which maps image features into the input space of language model, CogVLM bridges the gap between the frozen pretrained language model and image encoder by a trainable visual expert module in the attention and FFN layers. As a result, CogVLM enables deep fusion of vision language features without sacrificing any performance on NLP tasks. CogVLM-17B achieves state-of-the-art performance on 10 classic cross-modal benchmarks, including NoCaps, Flicker30k captioning, RefCOCO, RefCOCO+, RefCOCOg, Visual7W, GQA, ScienceQA, VizWiz VQA and TDIUC, and ranks the 2nd on VQAv2, OKVQA, TextVQA, COCO captioning, etc., surpassing or matching PaLI-X 55B. Codes and checkpoints are available at https://github.com/THUDM/CogVLM.