Text-Printed Image: Bridging the Image-Text Modality Gap for Text-centric Training of Large Vision-Language Models

Text-Printed Image: Bridging the Image-Text Modality Gap for Text-centric Training of Large Vision-Language Models Shojiro Yamabe1,2,† Futa Waseda1,3 Daiki Shiono1,4 Tsubasa Takahashi1,‡ 1Turing Inc. 2Institute of Science Tokyo 3The University of Tokyo 4Tohoku University †yamabe.s.2fb0@m.isct.ac.jp ‡tsubasa.takahashi@turing-motors.com Textual description “A man wearing a black suit and…” Text-printed image (ours) Q: Where is the man? A: At a media event. Text-to-Image Collecting real images is costly and limited Low Fidelity to Q&A Image-Text Modality Gap Language Model Image Encoder Higher Fidelity to Q&A Text-centric Training Text-centric Training GT-image t-SNE Visualization of Hidden States Modality Gap TPI bridges the modality gap Text TPI (ours) Image Encoder Text Encoder Figure 1. Text-Printed Image (TPI) provides an efficient and broadly applicable approach for text-centric training. While raw text input suffers from an image-text modality gap and synthetic images generated by a text-to-image model lose fidelity to Q&A pairs, TPI bridges this gap by embedding textual content into the visual pathway, achieving high fidelity to the ground-truth visual supervision. Abstract Recent large vision–language models (LVLMs) have been applied to diverse VQA tasks. However, achieving practi- cal performance typically requires task-specific fine-tuning with large numbers of image-text pairs, which are costly to collect. In this work, we study text-centric training, a set- ting where only textual descriptions are available and no real images are provided, as a paradigm for low-cost data scaling. Unlike images, whose collection is often restricted by privacy constraints and scarcity in niche domains, text is widely available. Moreover, text is easily editable, enabling automatic diversification and expansion with LLMs at min- imal human effort. While this offers clear advantages over image collection in terms of scalability and cost, training on raw text without images still yields limited gains on VQA tasks because of the image–text modality gap. To address this issue, we propose a Text-Printed Image (TPI), which generates synthetic images by directly rendering the given textual description on a plain white canvas. This simple ren- dering projects text into the image modality and can be in- tegrated into arbitrary existing LVLM training pipelines at low cost. Moreover, TPI preserves the semantics of the text, whereas text-to-image models often fail to do. Across four models and seven benchmarks, our systematic experiments show that TPI enables more effective text-centric training than synthetic images generated by a diffusion model. We further explore TPI as a low-cost data-augmentation strat- egy and demonstrate its practical utility. Overall, our find- ings highlight the significant potential of text-centric train- ing and, more broadly, chart a path toward fully automated data generation for LVLMs. 1. Introduction Large Vision-Language Models (LVLMs) [38] have emerged as a powerful extension of Large Language Mod- els (LLMs). Building on strong foundational capabilities, they have been applied across diverse VQA tasks, includ- 1 arXiv:2512.03463v1 [cs.CV] 3 Dec 2025 ing autonomous driving [72] and healthcare [21]. To at- tain practical performance on these downstream tasks, task- specific fine-tuning is typically required. However, constructing a large-scale dataset remains challenging for such training, as practical SFT typically re- quires vast amounts of image–text instructions [38, 61, 62]. In contrast to LLM training, which benefits from abundant text corpora [5], LVLM training requires image-conditioned instruction data, and constructing such data is substantially more costly [31, 38]. Although automated methods for scraping web images and pairing them with correspond- ing text have been proposed [10, 12, 30, 52, 67], their applicability is often limited in specialized or niche do- mains [32, 51]. In this work, we consider text-centric training, where only textual descriptions are available and no images are provided. Our premise is that text is abundant and easy to prepare compared to images. Rich text corpora and struc- tured lexical resources make it possible to synthesize de- scriptions from class names and concept taxonomies [16, 20, 58]. Moreover, text is much easier to modify than im- ages. While changing a small detail in an image is dif- ficult [22], textual descriptions can be edited or expanded with little effort, and LLMs can automatically generate di- verse variants [57]. Recent advances in LLM-based data augmentation [13, 56] further enhance this potential, en- abling scalable and diverse text synthesis at low cost. Taken together, these factors make text-centric training a promis- ing yet underexplored direction for advancing LVLMs when images are scarce or unavailable. A primary challenge for effective text-centric training is the image-text modality gap [17, 35, 66, 68], which is a

This content is AI-processed based on open access ArXiv data.