ExchangeDEX+

Buy Crypto Markets Spot FuturesGOLD Earn Event Center

Details 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.Details 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.

Visual Prompt Generation: Cross-Attention in Q-Former

2025/11/20 00:00

Table of Links

Abstract and 1 Introduction

Related Work

2.1. Multimodal Learning

2.2. Multiple Instance Learning
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
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
Conclusion and References

\ Supplementary Material

A. Detailed Architecture of QFormer

B. Proof of Proposition

C. More Experiments

\ Figure 7. Overview of QFormer

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].

:::info Authors:

(1) Wenliang Zhong, The University of Texas at Arlington (wxz9204@mavs.uta.edu);

(2) Wenyi Wu, Amazon (wenyiwu@amazon.com);

(3) Qi Li, Amazon (qlimz@amazon.com);

(4) Rob Barton, Amazon (rab@amazon.com);

(5) Boxin Du, Amazon (boxin@amazon.com);

(6) Shioulin Sam, Amazon (shioulin@amazon.com);

(7) Karim Bouyarmane, Amazon (bouykari@amazon.com);

(8) Ismail Tutar, Amazon (ismailt@amazon.com);

(9) Junzhou Huang, The University of Texas at Arlington (jzhuang@uta.edu).

:::

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

:::

Market Opportunity

Prompt Price(PROMPT)

$0.06551

$0.06551$0.06551

+2.50%

USD

Prompt (PROMPT) Live Price Chart

Disclaimer: The articles reposted on this site are sourced from public platforms and are provided for informational purposes only. They do not necessarily reflect the views of MEXC. All rights remain with the original authors. If you believe any content infringes on third-party rights, please contact service@support.mexc.com for removal. MEXC makes no guarantees regarding the accuracy, completeness, or timeliness of the content and is not responsible for any actions taken based on the information provided. The content does not constitute financial, legal, or other professional advice, nor should it be considered a recommendation or endorsement by MEXC.