InstructG2I: Synthesizing Images from Multimodal Attributed Graphs

In this paper, we approach an overlooked yet critical task Graph2Image: generating images from multimodal attributed graphs (MMAGs). This task poses significant challenges due to the explosion in graph size, dependencies among graph entities, and the need for controllability in graph conditions. To address these challenges, we propose a graph context-conditioned diffusion model called InstructG2I. InstructG2I first exploits the graph structure and multimodal information to conduct informative neighbor sampling by combining personalized page rank and re-ranking based on vision-language features. Then, a Graph-QFormer encoder adaptively encodes the graph nodes into an auxiliary set of graph prompts to guide the denoising process of diffusion. Finally, we propose graph classifier-free guidance, enabling controllable generation by varying the strength of graph guidance and multiple connected edges to a node. Extensive experiments conducted on three datasets from different domains demonstrate the effectiveness and controllability of our approach. The code is available at https://github.com/PeterGriffinJin/InstructG2I.

Paper

References (46)

Scroll for more · 34 remaining

Similar papers

Peer review

Reviewer 7dPX5/10 · confidence 3/52024-07-09

Summary

The authors propose an approach to enhance image synthesis using multimodal attributed graphs, adopting a strategy to condition image generation via a tokenization scheme on graph structure.

Strengths

- The paper studies an intersectional topic: leveraging graph learning techniques for image generation, which is a creative application and an area which deserves more focus. - The authors' use of qualitative examples (e.g. Figure 5 and 6) is commendable and helps articulate visual improvements.

Weaknesses

Please see questions and concerns below. My general feeling is the paper is fairly incremental in its introduction of a mechanism to encode graph condition into the conditioning for generation. Many design choices for graph conditioning are not discussed well and the quantitative results for some of these choices are missing which hurts the overall impact of the work.

Questions

- Typos: - Line 18: "graph-structued" - The motivation proposed in lines 28-30 is a little bit confusing, since the scenario the authors discuss here (e.g. virtual artwork creation based on nuanced styles of artists and genres) seems like it could be well-handled by text rather than explicitly using graph structure. - There is limited prior work in multimodal graph learning, as the authors mention. The authors may want to reference and position their work with respect to the recent [1] which offers multiple datasets and focuses on utility of GNN methods for node/link/graph-level tasks rather than generative tasks. - Nit: the notation is a bit awkward compared to conventional graph literature which typically uses $\mathcal{V}, \mathcal{E}, \mathcal{X}, \mathcal{F}$ or something similar to indicate node-set, edge-set, node-features, and edge-features. The authors proposed notation in line 72 seems to define P and D as different sets of images / documents compared to the nodes V, but then mentions that each node has some textual information and image information corresponding to P and D (it should be made clear whether this information is just features, or actual node relationships -- if the latter, it seems that P and D should be contained within the nodeset V). - The process described in line 115 around introducing tokens to help condition the representation using graph structure is also explored in some related works, e.g. [2]. Perhaps the authors could consider adopting a similar approach if it makes sense in this task, since the tokenization scheme as the authors of [2] point out is key in injecting the right level of information to the model. - Comment: the notation in pages 3-5 is quite heavy and would benefit from a symbol table. - Section 3.2 proposes a heuristic solution for neighbor selection. I'd encourage exploring solutions designed for learnable importance of multiple edge types similar to [3]. - Can the authors discuss what sort of techniques were used to incorporate graph conditions for the baseline models like InstructPix2Pix and SD? - Is there a quantitative understanding or experiment for the PPR neighbor based sampling approach? It seems this is one of the more heuristic parts of the paper where the design of the sampling procedure (two phase PPR + semantic re-ranking) is less conventional and deviates from other aggregation mechanisms explored in previous literature like attention-based selection, learnable importance of multiple edge types, etc. The qualitative experiment is helpful but not terribly convincing in terms of the actual performance impact in aggregate. [1] Multimodal Graph Benchmark (Zhu et al, 2024) [2] LLaGA: Large Language and Graph Assistant (Chen et al, ICML 2024) [3] Pathfinder Discovery Networks for Neural Message Passing (Rozemberczki et al, WWW 2021)

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes, Appendix A.1

Reviewer 7dPX2024-08-12

Thank you

Dear authors -- thank you for your response to my concerns. I think adding this discussion to the work will help strengthen the positioning (especially the parts around necessity/utility of graph conditioning vs. text). Overall, I stand by my initial review that parts of the work feel a bit heuristic and incremental, but the work is generally an interesting proposal which leans towards a different application of graph structure than most are used to seeing. I will retain my score.

Authorsrebuttal2024-08-13

Dear Reviewer 7dPX, Thank you so much for your reply! We will add those discussions to the revision according to your suggestions. We would like to emphasize the contributions of the work again here: - (**Problem**). We are pioneers in recognizing the potential of multimodal attributed graphs (MMAGs) for image synthesis and introducing the Graph2Image problem. - (**Algorithm**). We introduce InstructG2I, a context-aware diffusion model that adeptly encodes graph conditional information as graph prompts for controllable image generation. - (**Benchmark**). We construct a benchmark with graphs from three domains (art, e-commerce, literature) to evaluate the models on the Graph2Image problem. The benchmark can be used for future exploration of this problem. - (**Experiments**). We perform experiments on the benchmark, showing that InstructG2I consistently surpasses competitive baselines. For the (**Algorithm**) part, we propose four main novel components: - *Semantic PPR-based neighbor sampling*: It helps discover the informative neighbors based on semantic and structure information for target node image generation. - *Graph-QFormer*: It extracts the common features of the selected neighbors and aggregates the neighbor information as graph conditional tokens for the stable diffusion model, considering the image-image dependency and text-image dependency. - *Graph conditions as tokens*: It enables stable diffusion to leverage the graph condition information for target image generation. - *Graph-based controllable generation (Graph CFG)*: It enables controllable, tunable generation with multiple graph conditions. Although the philosophy of “graph tokens” is used in graph-enhanced text generation with LLMs, it is not explored in graph-enhanced image generation with diffusion models. Besides, it is nontrivial to get those graph tokens (semantic PPR-based neighbor sampling and Graph-QFormer) and use them for controllable generation (graph CFG). We would like to argue that even though the reviewer insists that “graph tokens” are incremental, it is only a small part of our **Algorithm** designs. We have other novel design components in **Algorithm** and other contributions in addition to **Algorithm** (**Problem**, **Benchmark**, **Experiments**). We would like to thank the reviewer again for the reply. If you have any more thoughts, we are happy to continue our discussion until the deadline.

Reviewer nWct6/10 · confidence 3/52024-07-12

Summary

This paper focuses on the problem of image synthesis on multimodal attributed graphs (MMAGs) and proposes a graph context-conditioned diffusion model, INSTRUCTG2I, to address the challenge in this setting. In particular, it proposes a semantic personalized PageRank-based method to sample related neighbors in the graph. Then, the INSTRUCTG2I can effectively encode graph conditional information as graph prompts with Graph-QFormer. Systematic experiments on MMAGs demonstrate the effectiveness of the methods proposed in this paper compared to competitive baseline methods.

Strengths

1. This paper studies an interesting and meaningful question. It investigates the graph-structured relationships of real-world entities for image generation on MMAGs, a task well-grounded in practical applications. 2. This paper is well-structured and easy to understand. 3. The graph context-conditioned diffusion model proposed in this paper is reasonable in solving image generation problems on MMAGs.

Weaknesses

1. The description in eq.10 may be incorrect. Please check more carefully. 2. Subsection 3.4 is more challenging to understand when reading. The authors' descriptions of some symbols in Eq. 10 and Eq. 11 are not exhaustive. 3. The results of the ablation experiments in Table 2 indicate that using a GNN such as GAT or GraphSAGE to aggregate graph information seems to be worse than the straightforward approach in Eq.7. Authors are requested to give a more detailed discussion with a reasonable explanation. 4. The images sampled by the semantic PPR-based sampling shown in Figure 5 appear to have the same image as the ground truth. Does this indicate that the proposed method suffers from label leakage?

Questions

1. Please see the weaknesses. 2. I wonder if the authors will compare it to other state-of-the-art image generation models, such as some Multimodal LLMs that are so prevalent nowadays.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

This paper has reasonably discussed the limitations.

Authorsrebuttal2024-08-11

Dear Reviewer nWct, We are glad that we address your raised issues and we are grateful that you increase the score! We will further enhance the submission according to your suggestions. Best, Authors

Reviewer 5RsF7/10 · confidence 3/52024-07-13

Summary

The paper introduces a new task graph2image which is to generate images conditioned on both text descriptions and graph information, which improves consistency of generated images compared to conditioned only on texts or images. To address combinatorial complexity of graphs and dependencies among graph entities, the paper proposes a graph context-conditioned diffusion model InstructG2I for generating images from multimodal attributed graph.

Strengths

- To the best of my knowledge, graph2image is a novel task, and the motivation to use the rich and high-dimensional information of graphs for image generation seems reasonable and interesting. - The proposed approach to incorporate graph information into pre-trained text-to-image is new, in particular introducing graph conditioning token and considering scalability of graph size. - The generated samples show that using graph information results in better consistency with the ground truth compared to methods that use only text prompts or images. - Examples of controllable generation with both text and graph show the ability to balance content and style in a simple manner.

Weaknesses

While I do not have a major concern, an ablation study on scalability to graph size seems to be missing. How large graphs is the method able to be applied?

Questions

- Why is the DINOv2 score on Goodreads dataset significantly low compared to that of ART500K or Amazon datsets?

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

4

Limitations

Yes the paper addresses the limitation in Appendix A.1.

Reviewer TQRx5/10 · confidence 3/52024-08-05

Summary

This paper introduces a novel approach for controllable image generation using both graph and text conditions. The authors propose that additional context information from multimodal attributed graphs (MMAGs) can enhance the performance of diffusion models. Specifically, they formulate the Graph2Image problem and develop the INSTRUCTG2I model to incorporate contextual information during the generation process. Empirical evaluations demonstrate the strong performance of the model.

Strengths

1. The paper is easy to follow. 2. The intuition behind the approach is clear.

Weaknesses

1. The overall setting is questionable. The authors integrate graph information using a Graph-QFormer and context information such as artists and genres, stored in graph prompt tokens. Given the large graph size, they only use subgraph structures. Consequently, the Stable Diffusion (SD) model absorbs additional information from similar artworks, which could be derived from image or text prompts alone. This raises the question of whether an additional condition structure is necessary. I suggest the authors demonstrate a unique application where standard models with text and image prompting capabilities are insufficient.

Questions

1. Are there any unique scenarios where only graph input can significantly improve SD performance? As my review is overdue, I welcome concise feedback and am open to clarifying any potential misunderstandings.

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

N/A

Reviewer nWct2024-08-11

Thank you for your detailed response, the author effectively addressed my issue, I will increase my score.

Area Chair C5pZ2024-08-11

reply to author's response

Dear reviewers, Could you take time to read the author's rebuttal and submit your feedback? Thank you. AC

Reviewer 5RsF2024-08-11

Thank you for the responses. I believe this paper tackles an interesting task with novel approach, and do not find any major concerns. Thus I will keep my score.

Authorsrebuttal2024-08-11

Dear Reviewer 5RsF, Thank you for your continuous support of our work! We will further enhance the submission according to your suggestions. Best, Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC