Visual Prompt Generation: Cross-Attention in Q-Former

Written by instancing | Published 2025/11/19
Tech Story Tags: deep-learning | cross-attention | q-former-architecture | bert-model | visual-prompt-embeddings | blip2 | multimodal-llms-(mllms) | learnable-queries

TLDRDetails the Q-Former architecture: a 12-layer BERT-based model using 32 learnable query embeddings. These queries use cross-attention to extract visual information for MLLM input.via the TL;DR App

Abstract and 1 Introduction

  1. Related Work

    2.1. Multimodal Learning

    2.2. Multiple Instance Learning

  2. Methodology

    3.1. Preliminaries and Notations

    3.2. Relations between Attention-based VPG and MIL

    3.3. MIVPG for Multiple Visual Inputs

    3.4. Unveiling Instance Correlation in MIVPG for Enhanced Multi-instance Scenarios

  3. Experiments and 4.1. General Setup

    4.2. Scenario 1: Samples with Single Image

    4.3. Scenario 2: Samples with Multiple Images, with Each Image as a General Embedding

    4.4. Scenario 3: Samples with Multiple Images, with Each Image Having Multiple Patches to be Considered and 4.5. Case Study

  4. Conclusion and References

Supplementary Material

A. Detailed Architecture of QFormer

B. Proof of Proposition

C. More Experiments

A. Detailed Architecture of QFormer

The architecture overview is depicted in Figure 7. Specifically, QFormer is initialized as a BERT-based model[8] comprising a total of L = 12 layers. In contrast to typical BERT models that process textual inputs, QFormer takes R = 32 learnable query embeddings as inputs. These embeddings are utilized to extract visual information from the input visual data during Stage-1 pretraining in BLIP2[22]. Subsequently, they serve as visual prompt embeddings for the LLM inputs after projection.

Inside the QFormer, each layer includes a self-attention module composed of a Multi-Head Attention component and a Forward module (consisting of Linear, LayerNorm, and Residual Connection). The cross-attention module, initialized with random values, is inserted every G layers, where learnable query embeddings interact with visual embeddings. In the main paper, for the sake of conciseness, we condensed the representation of the multi-head attention and forward modules into self(cross) attention modules. Furthermore, we exclusively illustrated the modifications made to the cross-attention module in MIVPG, as the self-attention modules remain unchanged. The final QFormer output is represented by the last layer’s query embeddings.

For a more comprehensive understanding, readers are encouraged to refer to [22].

Authors:

(1) Wenliang Zhong, The University of Texas at Arlington ([email protected]);

(2) Wenyi Wu, Amazon ([email protected]);

(3) Qi Li, Amazon ([email protected]);

(4) Rob Barton, Amazon ([email protected]);

(5) Boxin Du, Amazon ([email protected]);

(6) Shioulin Sam, Amazon ([email protected]);

(7) Karim Bouyarmane, Amazon ([email protected]);

(8) Ismail Tutar, Amazon ([email protected]);

(9) Junzhou Huang, The University of Texas at Arlington ([email protected]).


This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license.


Written by instancing | Pioneering instance management, driving innovative solutions for efficient resource utilization, and enabling a more sus
Published by HackerNoon on 2025/11/19