Generating images from textual arguments presents a significant challenge due to the inherent ambiguity and context-dependence of natural language. For the Touché 2025 shared task 'Image Retrieval/Generation for Arguments', we present a system that generates an image for an argument by first identifying and extracting key aspects of the argument. These aspects are then combined with the original argument to form a detailed prompt that describes a situation relevant to the argument. Finally, this prompt is used to generate a corresponding image. This approach allows us to guide image generation more efectively, ensuring that the output more accurately reflects the intended meaning of the argument. To evaluate the results, we employ several embedding methods to measure the semantic similarity between the arguments and the corresponding images. Our findings indicate that incorporating dedicated aspects into the image prompts significantly improves the quality and relevance of the generated images.
The objective of the shared task is to retrieve or generate images that efectively convey the central claims
of given arguments. Participants can either select images from a provided dataset of approximately
32,000 web-crawled images or employ an image generation model of their choice. The task comprises
128 arguments covering 27 distinct topics [
Identifying suitable images for textual arguments remains a significant challenge, as images are “ambiguous, yet rich in information” [5]. While images often convey more than words alone, their interpretation frequently relies on contextual cues. In previous iterations of the shared task, participants have explored how image generation can enhance retrieval — for instance, by producing reference images to compare against existing ones within a retrieval-based framework [6, 7]. In contrast, our approach is designed exclusively for image generation, without incorporating any retrieval component.
Initial experiments using the raw argument text as a prompt for image generation revealed a recurring issue: many of the resulting images failed to capture core aspects of the input arguments. Instead of conveying the intended semantic content, the generated images often included irrelevant or distracting visual elements, introducing noise that compromised both their interpretability and overall efectiveness. To address this limitation, our approach focuses on visualizing the key aspects of each argument. This is achieved through the integration of several dedicated modules. An overview of our system architecture is shown in Figure 1.
Argument
Aspects
Image Prompt Image Generator
Is Image good? (Human Evaluation)
Final Image
The first step in our system is to identify which aspects of the argument should be visualized in the corresponding image. To do this, we use the LLaMA 3.2 (3B-Instruct) language model [8, 9] to extract three key aspects that capture the most important elements of the argument. These aspects serve as essential components that must be visually represented in the generated image. In the second step, the list of extracted aspects, together with the original argument, is passed to the argument-image prompt generator module. This module leverages the Mistral (7B-v0.1) model [10, 11] to create a detailed prompt for the image generator (image prompt). Mistral was chosen for this task due to its superior performance compared to other large language models [10]. The image prompt describes the argument as a vivid scene, carefully incorporating each identified aspect. Table 1 demonstrates how the final image prompt difers from the original argument. In the third step, the image prompt is utilized to generate
Image Prompt: A man sitting in a fast food restaurant, surrounded by empty wrappers and packaging. He has a large belly, and his face is red and sweaty from overeating. the corresponding image. We employ Stable Difusion XL (Base 1.0) [ 12, 13] as the image generation model, configured with a fixed seed value of 1244 and 40 inference steps to ensure consistency and reproducibility. This model is selected over alternative models due to its capability for producing softer, more artistic renderings — particularly well-suited for stylized, anime, and fantasy art — as evidenced by the comparative results presented in Table 1. The fourth step involves a quality check performed by a human expert, who evaluates whether the generated image efectively represents the key aspects of the argument. If the image is deemed satisfactory, it is accepted; otherwise, the prompt is revised to emphasize any missing elements. However, for the images submitted to the task, no prompt revisions were required. Using this process, five images were generated for each argument.
Finally, the generated images are ranked by first generating a description of each image using LLaVA 1.5 (13B) [14, 15], and then computing the cosine similarity between this description and the original image prompt using Sentence-BERT (SBERT) [16], specifically the all-MiniLM-L6-v2 model. An example of this similarity assessment and ranking method is illustrated in Table 2 in the appendix.
To evaluate how well the generated images align with the arguments, we employ three distinct methods to measure semantic similarity between argument-image pairs. The first method directly compares the image and the argument by computing cosine similarity between their embeddings, generated using the multimodal CLIP model (clip-vit-base-patch32) [17]. For the second and third methods, we ifrst generate a description of each image using LLaVA 1.5 (13B) [ 14, 15]. In the second method, the image description is compared to the corresponding argument by computing SBERT embeddings and measuring their cosine similarity. The third method compares the image description to the image prompt used to generate the image, also using SBERT embeddings. This third method is also employed internally by the system to determine the final ranking of the generated images, as detailed in Section 3. A summary of the mean cosine similarity scores for all three methods across individual topics, along Image - Argument Image Description - Image Prompt Image Description - Argument 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Topic Number Evaluation
Image Description - Argument 0.3266 with central metrics averaged across all approaches, is presented in Figure 2. More detailed results for each of the three evaluation methods are available in the appendix, as shown in Figure 3. The results show that the cosine similarity between the image description generated by LLaVA and the image prompt is the highest among all measured similarities, consistently exceeding the similarity between the image description and the original argument. This is likely because the image prompts contain more detailed descriptions of visual aspects than the arguments alone, leading to higher similarity scores. Overall, the consistently high similarity confirms that the generated images closely align with the image prompts. In addition, we observe a strong correlation between similarities based on the image prompt and those based on the argument. This indicates that the image prompt is thematically aligned with the argument. Some topics in Figure 2 exhibit noticeably lower similarity. These cases may correspond to complex arguments, or to arguments that require the visualization of undesirable or negative aspects.
The CLIP-based similarity between the image and the original argument consistently produces the lowest similarity scores, with the values lying very close together. This observation aligns with existing research [18], which highlights that CLIP similarities are generally very uniform. This insight plays a particularly important role in automatically deciding whether an image fits an argument. In such cases, CLIP embeddings are likely not suitable.
Notably, as shown in Table 2 in the appendix, the image prompts used to guide generation difer substantially from the original arguments. Through this transformation and fusion, the visual essence of the argument is better captured, resulting in improved image quality, enhanced visual clarity, and reduced visual noise. The strength of our approach for generating highly relevant images is demonstrated by its first-place ranking in the shared task, outperforming all baseline approaches that rely on the raw argument text as prompts.
In this work, we introduced a novel argument-image generation system that first identifies relevant aspects of the arguments and then combines these aspects with the original argument to create a corresponding image prompt. This prompt is used by an image generator to produce the related image. By enriching the original arguments with key aspects and using detailed, extended prompts for generation, our approach improves image quality and reduces visual noise. This enables the generation of highly relevant images and resulted in the system achieving first place in the competition.
This work was partially supported by the German Federal Ministry of Education and Research (BMBF) through the project “DIALOKIA: Überprüfung von LLM-generierter Argumentation mittels dialektischem Sprachmodell” (01IS24084A-B).
During the preparation of this work, the authors used ChatGPT, DeepL, Grammarly, Grok and Language Tool in order to: grammar and spelling check, paraphrase and reword, improve writing style. The presented image generation pipeline integrates several AI models: prompts are produced with LLaMA, Stable Difusion is used to generate images, and LLaVA provides automated image descriptions. After using these tools/services, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. in Information Retrieval. 45th European Conference on IR Research (ECIR 2023), Lecture Notes in Computer Science, Springer, Berlin Heidelberg New York, 2023, pp. 236–241. doi:10.1007/ 978-3-031-28241-6_20. [5] J. E. Kjeldsen, The rhetoric of thick representation: How pictures render the importance and strength of an argument salient, Argumentation 29 (2015) 197–215. URL: https://doi.org/10.1007/ s10503-014-9342-2. doi:10.1007/s10503-014-9342-2. [6] B. Ostrower, P. Aphiwetsa, Ds@gt at touché: Image search and ranking via clip and image generation, in: G. Faggioli, N. Ferro, P. Galuscakova, A. G. S. Herrera (Eds.), Working Notes Papers of the CLEF 2024 Evaluation Labs, volume 3740 of CEUR Workshop Proceedings, 2024, pp. 3447–3450. URL: http://ceur-ws.org/Vol-3740/paper-332.pdf. [7] M. Moebius, M. Enderling, S. T. Bachinger, Jean-luc picard at touché 2023: Comparing image generation, stance detection and feature matching for image retrieval, in: M. Aliannejadi, G. Faggioli, N. Ferro, M. Vlachos (Eds.), Working Notes Papers of the CLEF 2023 Evaluation Labs, volume 3497 of CEUR Workshop Proceedings, 2023, pp. 3111–3118. URL: http://ceur-ws.org/Vol-3497/paper-263.pdf. [8] A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fan, A. Goyal, A. Hartshorn, A. Yang, A. Mitra, A. Sravankumar, A. Korenev, A. Hinsvark, A. Rao, A. Zhang, A. Rodriguez, A. Gregerson, A. Spataru, B. Rozière, B. Biron, B. Tang, B. Chern, C. Caucheteux, C. Nayak, C. Bi, C. Marra, C. McConnell, C. Keller, C. Touret, C. Wu, C. Wong, C. C. Ferrer, C. Nikolaidis, D. Allonsius, D. Song, D. Pintz, D. Livshits, D. Esiobu, D. Choudhary, D. Mahajan, D. Garcia-Olano, D. Perino, D. Hupkes, E. Lakomkin, E. AlBadawy, E. Lobanova, E. Dinan, E. M. Smith, F. Radenovic, F. Zhang, G. Synnaeve, G. Lee, G. L. Anderson, G. Nail, G. Mialon, G. Pang, G. Cucurell, H. Nguyen, H. Korevaar, H. Xu, H. Touvron, I. Zarov, I. A. Ibarra, I. M. Kloumann, I. Misra, I. Evtimov, J. Copet, J. Lee, J. Gefert, J. Vranes, J. Park, J. Mahadeokar, J. Shah, J. van der Linde, J. Billock, J. Hong, J. Lee, J. Fu, J. Chi, J. Huang, J. Liu, J. Wang, J. Yu, J. Bitton, J. Spisak, J. Park, J. Rocca, J. Johnstun, J. Saxe, J. Jia, K. V. Alwala, K. Upasani, K. Plawiak, K. Li, K. Heafield, K. Stone, et al., The llama 3 herd of models, CoRR abs/2407.21783 (2024). URL: https: //doi.org/10.48550/arXiv.2407.21783. doi:10.48550/ARXIV.2407.21783. arXiv:2407.21783. [9] Meta AI, meta-llama/llama-3.2-3b-instruct model card, https://huggingface.co/meta-llama/Llama-3.
2-3B-Instruct, 2024. Model card accessed in June 2025. [10] A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. de Las Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, L. R. Lavaud, M. Lachaux, P. Stock, T. L. Scao, T. Lavril, T. Wang, T. Lacroix, W. E. Sayed, Mistral 7b, CoRR abs/2310.06825 (2023). URL: https://doi.org/10.48550/ arXiv.2310.06825. doi:10.48550/ARXIV.2310.06825. arXiv:2310.06825. [11] Mistral AI Team, mistralai/mistral-7b-v0.1 model card, https://huggingface.co/mistralai/
Mistral-7B-v0.1, 2023. Hugging Face model card accessed June 2025. [12] D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, R. Rombach, SDXL: improving latent difusion models for high-resolution image synthesis, in: The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, OpenReview.net, 2024. URL: https://openreview.net/forum?id=di52zR8xgf. [13] Stability AI, stabilityai/stable-difusion-xl-base-1.0 Model Card, https://huggingface.co/stabilityai/ stable-difusion-xl-base-1.0, 2023. Model card accessed June 2025; includes details on ensemble pipeline, text encoders, and license (OpenRAIL++-M). [14] C. Li, C. Wong, S. Zhang, N. Usuyama, H. Liu, J. Yang, T. Naumann, H. Poon, J. Gao, Llava-med: Training a large language-and-vision assistant for biomedicine in one day, in: A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, S. Levine (Eds.), Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. URL: http://papers.nips.cc/paper_files/paper/ 2023/hash/5abcdf8ecdcacba028c6662789194572-Abstract-Datasets_and_Benchmarks.html. [15] LLaVA Contributors, user/llava-1.5-13b-hf model card, https://huggingface.co/user/llava-1. 5-13b-hf, 2025. Hugging Face model card accessed June 2025; variant of the LLaVA-1.5 13 B model with HF integration. [16] N. Reimers, I. Gurevych, Sentence-bert: Sentence embeddings using siamese bert-networks, in: K. Inui, J. Jiang, V. Ng, X. Wan (Eds.), Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, Association for Computational Linguistics, 2019, pp. 3980–3990. URL: https://doi.org/10.18653/v1/D19-1410. doi:10.18653/V1/D19-1410. [17] OpenAI, openai/clip-vit-base-patch32 model card, https://huggingface.co/openai/ clip-vit-base-patch32, 2021. Model card accessed June 2025, covers architecture, intended use, performance, and biases. [18] K. Tyshchuk, P. Karpikova, A. Spiridonov, A. Prutianova, A. Razzhigaev, A. Panchenko, On isotropy of multimodal embeddings, Information-an International Interdisciplinary Journal 14 (2023). doi:10.3390/info14070392.
Source Image - Argument Image Description - Image Prompt Image Description - Argument Overal 0.0 0.2 0.4 Cosine Similarity 0.6 0.8 A young adult woman is sitting at a table with a plate of french fries, a burger, and a soda in front of her. She is looking down at the food and is frowning, clearly feeling guilty about her unhealthy eating habits. She is wearing jeans and a t-shirt, and her hair is pulled back in a ponytail. Behind her, there is a
restaurant, surrounded by empty wrappers and packaging. He has a large belly, and his face is red and sweaty from overeating.
ting at a dining table with a plate of food in front of her. The plate contains a hamburger, french fries, and a sandwich. She is holding a sandwich in her hand, and there is a cup on the table as well. The woman appears to be enjoying her meal, and the scene is set in a restaurant.
ing of a man sitting on a chair.
He is wearing a white shirt and red pants. The man appears to be quite large, possibly obese. He is surrounded by a large number of bags of chips, with some of them scattered on the floor. The man seems to be enjoying his time, possibly eating the chips.
A group of children playing The image depicts a group in a park, surrounded by fast of children gathered around food containers, chips, and a tree, enjoying a picnic tosugary drinks. The children gether. They are sitting on the are oblivious to the unhealthy ground, with some of them consequences of their snack- eating chips and drinking from ing habits, and continue to eat cups. There are several cups unhealthy food, ignoring the scattered around the area, warnings from their parents along with a few bottles. The and health experts. children are engaged in conversation and laughter, creating a lively and fun atmosphere.
CS
1 2 3