More Works from Our Lab

Emuru: Zero-Shot Styled Text Image Generation, but Make It Autoregressive

First autoregressive model for zero-shot styled text image generation with generalization to novel styles.

CVPR 2025
Merging and Splitting Diffusion Paths for Semantically Coherent Panoramas

Efficient generation of panoramic images with desired characteristics by adapting diffusion models.

ECCV 2024
Alfie: Democratising RGBA Image Generation With No $$$

Lightweight generation of RGBA images with transparency using efficient diffusion model adaptation.

ECCVW 2024
Show More Works

Autoregressive Styled Text Image Generation, but Make it Reliable

Carmine Zaccagnino1, Fabio Quattrini1, Vittorio Pippi1, Silvia Cascianelli1, Alessio Tonioni2, Rita Cucchiara1
1University of Modena and Reggio Emilia    2Google
WACV 2026
Paper Code 🤗 Demo 🤗 PT Dataset 🤗 FT Dataset arXiv Cite
Eruku generates text images with arbitrary length and great text adherence

Eruku can generate text images with arbitrary length and great text adherence, respecting both the generation text and the conditioning writing style. Our approach improves upon previous autoregressive methods by making style text optional and introducing explicit stopping mechanisms.

Abstract

Generating faithful and readable styled text images (especially for Styled Handwritten Text generation - HTG) is an open problem with several possible applications across graphic design, document understanding, and image editing. A lot of research effort in this task is dedicated to developing strategies that reproduce the stylistic characteristics of a given writer, with promising results in terms of style fidelity and generalization achieved by the recently proposed Autoregressive Transformer paradigm for HTG. However, this method requires additional inputs, lacks a proper stop mechanism, and might end up in repetition loops, generating visual artifacts. In this work, we rethink the autoregressive formulation by framing HTG as a multimodal prompt-conditioned generation task, and tackle the content controllability issues by introducing special textual input tokens for better alignment with the visual ones. Moreover, we devise a Classifier-Free-Guidance-based strategy for our autoregressive model. Through extensive experimental validation, we demonstrate that our approach, dubbed Eruku, compared to previous solutions requires fewer inputs, generalizes better to unseen styles, and follows more faithfully the textual prompt, improving content adherence.

Method

Eruku architecture overview

Eruku Training Framework. We condition generation on the textual content of a style image, the generation text, and the style image itself. Eruku learns to generate images containing the generation text with the same writing style as the style image. The model can work without style text by using synchronization tokens to separate sequence components.

Key Features

🎯 Optional Style Text

Unlike previous methods, Eruku can generate styled text images without requiring the transcription of the style image, making it more practical for real-world applications.

🛑 Explicit Stopping

Introduces a dedicated end-of-generation token instead of relying on heuristics, improving generation efficiency and reliability.

✨ CFG-Inspired Strategy

Employs Classifier-Free Guidance on textual inputs to enforce better adherence to the desired text sequence without auxiliary networks.

Results

Eruku generation results on various datasets

Qualitative Results. Eruku generates high-quality styled text images across different datasets (IAM, CVL, RIMES) with accurate text content and faithful style reproduction. The model works on both handwritten and typewritten styles.

Quantitative Comparison

Eruku demonstrates significant improvements over the previous state-of-the-art autoregressive method (Emuru):

  • Better Content Adherence: Lower ΔCER (Character Error Rate difference) showing more faithful text generation
  • Improved Style Fidelity: Lower HWD (Handwriting Distance) indicating better style reproduction
  • Fewer Required Inputs: Can operate without style text transcription
  • More Robust: Better performance even with noisy OCR-generated style text

BibTeX

@inproceedings{zaccagnino2026eruku,
  title={{Autoregressive Styled Text Image Generation, but Make it Reliable}},
  author={Zaccagnino, Carmine and Quattrini, Fabio and Pippi, Vittorio and Cascianelli, Silvia and Tonioni, Alessio and Cucchiara, Rita},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year={2026}
}