Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ovis-Image Technical Report

Guo-Hua Wang, Liangfu Cao, Tianyu Cui, Minghao Fu, Xiaohao Chen, Pengxin Zhan, Jianshan Zhao, Lan Li, Bowen Fu, Jiaqi Liu, Qing-Guo Chen

TL;DR

Ovis-Image delivers a 7B text-to-image model optimized for high-quality in-image text rendering under practical compute limits by pairing a diffusion-based visual decoder with a strong Ovis 2.5 multimodal backbone. The approach uses a four-stage training pipeline (pretraining, SFT, DPO, GRPO) and carefully curated data (including bilingual and synthetic text-rendering samples) to achieve frontier-like text rendering while maintaining general image quality and efficiency. Empirical results across CVTG-2K, LongText-Bench, DPG-Bench, GenEval, and OneIG-Bench show competitive or superior text rendering and bilingual capabilities relative to larger open and closed models, with favorable computational overhead. The work highlights a design principle: explicit, text-centric alignment and data curation can yield substantial text rendering gains without resorting to tens-of-billions-parameter models, enabling practical deployment on single high-end GPUs.

Abstract

We introduce $\textbf{Ovis-Image}$, a 7B text-to-image model specifically optimized for high-quality text rendering, designed to operate efficiently under stringent computational constraints. Built upon our previous Ovis-U1 framework, Ovis-Image integrates a diffusion-based visual decoder with the stronger Ovis 2.5 multimodal backbone, leveraging a text-centric training pipeline that combines large-scale pre-training with carefully tailored post-training refinements. Despite its compact architecture, Ovis-Image achieves text rendering performance on par with significantly larger open models such as Qwen-Image and approaches closed-source systems like Seedream and GPT4o. Crucially, the model remains deployable on a single high-end GPU with moderate memory, narrowing the gap between frontier-level text rendering and practical deployment. Our results indicate that combining a strong multimodal backbone with a carefully designed, text-focused training recipe is sufficient to achieve reliable bilingual text rendering without resorting to oversized or proprietary models.

Ovis-Image Technical Report

TL;DR

Ovis-Image delivers a 7B text-to-image model optimized for high-quality in-image text rendering under practical compute limits by pairing a diffusion-based visual decoder with a strong Ovis 2.5 multimodal backbone. The approach uses a four-stage training pipeline (pretraining, SFT, DPO, GRPO) and carefully curated data (including bilingual and synthetic text-rendering samples) to achieve frontier-like text rendering while maintaining general image quality and efficiency. Empirical results across CVTG-2K, LongText-Bench, DPG-Bench, GenEval, and OneIG-Bench show competitive or superior text rendering and bilingual capabilities relative to larger open and closed models, with favorable computational overhead. The work highlights a design principle: explicit, text-centric alignment and data curation can yield substantial text rendering gains without resorting to tens-of-billions-parameter models, enabling practical deployment on single high-end GPUs.

Abstract

We introduce , a 7B text-to-image model specifically optimized for high-quality text rendering, designed to operate efficiently under stringent computational constraints. Built upon our previous Ovis-U1 framework, Ovis-Image integrates a diffusion-based visual decoder with the stronger Ovis 2.5 multimodal backbone, leveraging a text-centric training pipeline that combines large-scale pre-training with carefully tailored post-training refinements. Despite its compact architecture, Ovis-Image achieves text rendering performance on par with significantly larger open models such as Qwen-Image and approaches closed-source systems like Seedream and GPT4o. Crucially, the model remains deployable on a single high-end GPU with moderate memory, narrowing the gap between frontier-level text rendering and practical deployment. Our results indicate that combining a strong multimodal backbone with a carefully designed, text-focused training recipe is sufficient to achieve reliable bilingual text rendering without resorting to oversized or proprietary models.

Paper Structure

This paper contains 19 sections, 6 equations, 2 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Architecture
  3. Data Composition
  4. Data for Pre-training
  5. Data for Supervised Fine-tuning
  6. Data for DPO
  7. Data for GRPO
  8. Training
  9. Training Infrastructure
  10. Training Procedure
  11. Evaluation
  12. CVTG-2K
  13. LongText-Bench
  14. DPG-Bench
  15. GenEval
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Comprehensive illustration of the functional capabilities of Ovis-Image.
  • Figure 2: The overall architecture of Ovis-Image. The architecture of Ovis-Image builds upon Ovis-U1, enhancing its capabilities by increasing the parameters of MMDiT and streamlining the structural design to create a more efficient and refined overall framework.