How TextFlux Enables OCR‑Free Multi‑Language Scene Text Editing with Diffusion Models

Introduction

In the field of text editing, TextFlux proposes an OCR‑free technique that embeds text into real‑world scenes using only glyph images and a small amount of training data, supporting Chinese, Japanese, Korean and other languages.

Related Links

Code: https://huggingface.co/yyyyyxie/textflux

Weights: https://huggingface.co/yyyyyxie/textflux

Demo page: https://yyyyyxie.github.io/textflux-site/

ComfyUI integration: https://github.com/yyyyyxie/textflux_comfyui

Background

Scene text editing must balance spelling accuracy and natural visual integration. Traditional methods either produce a patchy effect or fail to recognize text. By leveraging the contextual reasoning of diffusion transformers (DiT), TextFlux discards OCR encoders and directly fuses rendered glyph images with scene images, letting the model focus on seamless visual blending.

Multi‑line Generation

TextFlux can generate multiple lines of text in a single pass, allowing precise control over content, position, and style while maintaining readability and realism.

Key Features (Multi‑line)

OCR‑Free architecture: no OCR encoder required.

High fidelity and consistent style with the surrounding scene.

Multilingual support with low‑resource adaptation (under 1,000 samples for new languages).

Zero‑shot generalization to unseen characters.

Controllable multi‑line synthesis with per‑line layout control.

Data‑efficient training using only ~1% of data compared to other methods.

Single‑line Generation

To improve quality for single‑line scenarios, a beta model uses a line‑image condition instead of full‑size masks, reducing computation and providing more stable supervision, resulting in higher‑quality rendering for small fonts.

Key Features (Single‑line)

Significant improvement in single‑line rendering quality.

~1.4× faster inference for single lines.

Much higher accuracy for small‑size text synthesis.

Results

Outstanding Multilingual Editing

TextFlux excels at multi‑language scene text editing, generating high‑quality text even for low‑resource languages with fewer than 1,000 samples, dramatically lowering annotation costs.

Visual Fidelity and Text Accuracy

Both quantitative and qualitative evaluations show TextFlux surpasses state‑of‑the‑art methods in visual realism and textual accuracy, especially on low‑resolution inputs, and remains competitive with much larger LLM‑based image generation models.

Comparison with Diffusion‑Based Methods

Comparison with LLM‑Based Text‑to‑Image Models

Scalability and Zero‑Shot Capability

Thanks to its OCR‑Free architecture, TextFlux can instantly render characters unseen during training (zero‑shot), demonstrating strong generalization to rare Chinese characters and low‑resource scripts, though handwritten cursive scripts like Arabic remain challenging.

Preserving Base Model Generality

TextFlux does not alter the underlying base model's general capabilities, allowing the provided weights to extend other models built on the same foundation with multilingual generation abilities.

Conclusion

TextFlux breaks the bottleneck of scene text editing by introducing an OCR‑Free architecture that guides diffusion transformers with glyph images, achieving a compact structure, high data efficiency, multilingual and multi‑line support, and zero‑shot generalization, thereby improving visual realism and text accuracy for practical applications.