What Is Learned by DreamLLM? Dream Query Attention

cover
27 Nov 2024

Abstract and 1 Introduction

2 Background & Problem Statement

2.1 How can we use MLLMs for Diffusion Synthesis that Synergizes both sides?

3 DreamLLM

3.1 End-to-End Interleaved generative Pretraining (I-GPT)

3.2 Model Training

4 Experiments and 4.1 Multimodal Comprehension

4.2 Text-Conditional Image Synthesis

4.3 Multimodal Joint Creation & Comprehension

5 Discussions

5.1 Synergy between creation & Comprehension?

5. 2 What is learned by DreamLLM?

6 Related Works

7 Conclusions and References

A Additional Experiments

B Additional Qualitative Examples

C Implementation Details

D Additional Related Works

E Limitations, Failure Cases & Future Works

5.2 WHAT IS LEARNED BY DREAMLLM?

Dream Query Attention In DREAMLLM, the conditional embedding is derived from MLLMs with some learned dream queries. Fig. 6 demonstrates a visualization of the learned cross-attention mechanism between these queries and the diffusion latent. Similar to (Hertz et al., 2023), we visualize the attention map averaged across all timestamps. It is seen that: i) The query attention is structured, disentangled, and semantically-oriented.

This is evidenced by the fact that distinct queries adeptly capture different subject and background semantics. ii) Despite varying prompts, attention patterns exhibit remarkable similarity as shown in Fig. 6 (a) and (b). This contrasts with the token attentions from the original SD, which are typically text-token dependent. We postulate that this arises from the model’s causal nature, leading to a consistent semantic structure order.

Figure 6: Cross-attention of dream queries and the diffusion U-Net latent. Similar to (Hertz et al., 2023), the 64 queries can be viewed as 64 “words”. Each attention map is computed as the cross-attention between each query and the latent feature in the U-Net. The 64 queries are ordered as 8×8 grid sequentially, and each attention map is the result averaged across all timestamps.

This paper is available on arxiv under CC BY-NC-ND 4.0 DEED license.

Authors:

(1) Runpei Dong, Xi’an Jiaotong University and Internship at MEGVII;

(2) Chunrui Han, MEGVII Technology;

(3) Yuang Peng, Tsinghua University and Internship at MEGVII;

(4) Zekun Qi, Xi’an Jiaotong University and Internship at MEGVII;

(5) Zheng Ge, MEGVII Technology;

(6) Jinrong Yang, HUST and Internship at MEGVII;

(7) Liang Zhao, MEGVII Technology;

(8) Jianjian Sun, MEGVII Technology;

(9) Hongyu Zhou, MEGVII Technology;

(10) Haoran Wei, MEGVII Technology;

(11) Xiangwen Kong, MEGVII Technology;

(12) Xiangyu Zhang, MEGVII Technology and a Project leader;

(13) Kaisheng Ma, Tsinghua University and a Corresponding author;

(14) Li Yi, Tsinghua University, a Corresponding authors and Project leader.