Research map / Vision-language and multimodal / Multimodal LLMs

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

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Text-Printed Image (TPI), which renders text descriptions onto a white canvas as training images so LVLMs can be trained text-centrically with less modality gap than raw text or diffusion-generated images.

Problem

Text-centric training of large vision-language models, where only textual descriptions exist and no real images are available, suffers from an image-text modality gap, and diffusion-generated substitute images often fail to match the text.

Approach

Render each textual description onto a plain white canvas with a deterministic renderer and feed it through the unmodified image pathway (vision encoder and projector) during LoRA fine-tuning, so text supervision passes through the visual modality without architectural changes or inference-time modifications.

Key results

• Averaged over seven tasks, TPI reaches 49.97 for LLaVA 7B versus 47.12 for text-only and 48.43 for T2I, against a GT-image oracle of 55.58.
• On Qwen2.5-VL, TPI averages 77.94 versus 72.47 for T2I, and on LLaMA Vision it averages 72.27 versus 60.36 for T2I.
• TPI image generation runs at about 154 images/s on CPU versus 0.16 images/s for SDXL on one H100, and the relevance score averages 63.61 (90.8% of GT) versus 32.45 (46.3%) for T2I.

Surprising findings

• Text-only training on Qwen2.5-VL performs close to GT-image training on several tasks, yet shows much larger drift in intermediate representations than TPI.
• Font size, font color, and description-generation prompt have little effect on TPI performance, suggesting semantic content matters more than typography.

Limitations

• TPI assumes the LVLM already has visual understanding; it cannot teach new visual concepts such as shape or color from text alone.
• TPI requires a minimum level of OCR ability, so gains are smaller for weaker-OCR models such as LLaVA 7B/13B.
• Experiments use one generation pipeline with Qwen2.5-VL-32B descriptions derived from GT images and a minimal Self-Instruct augmentation.

Experimental setup

Datasets: ScienceQA OK-VQA VizWiz ChartQA InfoVQA DocVQA DriveLM

Benchmarks: ScienceQA OK-VQA VizWiz ChartQA InfoVQA DocVQA DriveLM OCRBench TextVQA

Metrics: accuracy JS divergence CKA relevance score gap ratio throughput

Baselines: Text-only training Text-to-Image (Stable Diffusion XL) GT-Image oracle

Builds on: LLaVA Self-Instruct VQAScore LoRA

Closest papers by meaning

• The Narrow Gate: Localized Image-Text Communication in Native Multimodal Models (NeurIPS 2025 · similarity 0.90)
• Mostly Text, Smart Visuals: Asymmetric Text-Visual Pruning for Large Vision-Language Models (CVPR 2026 · similarity 0.90)
• Words or Vision: Do Vision-Language Models Have Blind Faith in Text? (CVPR 2025 · similarity 0.89)
• MARS: Mixture of Auto-Regressive Models for Fine-grained Text-to-image Synthesis (AAAI 2025 · similarity 0.89)
• MANZANO: A Simple and Scalable Unified Multimodal Model with a Hybrid Vision Tokenizer (ICLR 2026 · similarity 0.89)
• LensVLM: Selective Context Expansion for Compressed Visual Representation of Text (NeurIPS 2026 · similarity 0.89)