FLUX-Text is an open source diffusion transformer model for scene text editing (Scene Text Editing). It supports multi-lingual text modification, addition and repair. It is suitable for image and text processing scenarios such as poster editing, emoticon production, and advertising design.

FLUX-Text

FLUX-Text (Flux Text) is a multilingual scene text editing model (Scene Text Editing, STE) open sourced by the AMAP machine learning team. It focuses on modifying, erasing, or adding new text to existing text in images. Different from general-purpose graphic models (such as Stable Diffusion and Midjourney), the core capability of FLUX-Text is to "precisely control the text content in the image" - users can specify to change the Chinese title on the poster to English, modify the text bubble text in the emoticon package, or replace the price text on the product image - and the modification results maintain natural integration visually, leaving no traces of editing.

Core parameters and statistics

Parameter items	Specifications
Product name	FLUX-Text
Category	AI Image Editing/Scene Text Editing
Delivery form	Open source model (pre-trained weights + inference code)
Open source license	Subject to the official repository
Model Architecture	Diffusion Transformer (DiT)
Supported languages	Chinese, English and other languages
Core tasks	Scene text editing (modify/add/erase text in images)
Project Ownership	AMAP ML Team (Alibaba / AMAP)
Paper published	arXiv (2025-05)
Open Source Platform	GitHub + Hugging Face
Demo environment	Gradio Demo
Research team	Rui Lan et al

The fundamental difference between FLUX-Text and general-purpose text drawing models: P-drawing software and general-purpose AI image tools can usually only do "whole image redrawing" when processing text in images, and cannot precisely control the modification of specific text areas - they may change the background, characters, or composition. FLUX-Text only operates on the specified text area, leaving other parts of the image unchanged. This makes it irreplaceable in scenarios such as poster editing, advertising modification, expression pack production, product image price update, etc. that require "only changing the words but not the image".

User and market recognition

As an academic open source project, FLUX-Text's influence is mainly reflected in the research community and open source community. The paper was published on arXiv (2025-05) and has received certain attention in the subdivided research field of scene text editing (STE). The GitHub repository provides complete inference code and pre-trained weights, and the Gradio Demo hosted on Hugging Face allows users to experience the model effects online.

At present, FLUX-Text has not yet formed a broad end-user market - its main audience is AI researchers and developers, rather than directly targeting C-end consumers. The product form is an open source model rather than a SaaS platform, which means that users need to have certain technical capabilities (Python environment configuration, model download and inference) to use it. The core indicators of market recognition - the number of paper citations, GitHub Stars, and Hugging Face downloads - are recommended to be based on the real-time data of each platform, and no speculative quotations will be made here.

Cost advantage

Cost Dimension	Description
Software License	$0 (open source model, subject to official license)
Inference computing	GPU required (16GB+ VRAM recommended), cloud GPU rental is about $0.5-2/hour
Deployment and operation	Requires Python and PyTorch environment configuration
API Service	No official hosted API; Gradio Demo is for research and demonstration purposes

The core cost advantages of the open source model: zero license fees + fully customizable. Compared with commercial image editing SaaS (such as Adobe Firefly, Canva AI) based on subscription ($10-60/month), FLUX-Text is suitable for teams with technical capabilities for private deployment and batch processing. For an e-commerce scenario that needs to process text replacement for 1,000 product images every day, the marginal cost of self-deploying FLUX-Text is mainly GPU computing power (about $15-30/day), while the cost of bulk purchasing commercial AI editing services is usually $100-500/day.

Hidden costs: Model inference requires GPU resources (16GB+ VRAM recommended), which cannot be used directly by non-technical users; the model may not be ideal for editing unconventional fonts (such as handwriting, word art); the current version does not support scene text editing in videos.

Main functions

• Scene text modification: Replace the text content in the specified area of ​​the image with the target text. For example: change the "2025 Spring Conference" in the poster to "2025 Summer Promotion", and the model automatically matches the original text font style, font size, color and perspective angle, making the modified text seamlessly blend into the background visually.
• Scene Text Erase: Erase unwanted text areas in the image and fill the erased areas with reasonable background content. For example: remove brand logo text from product images, and remove timestamps from screenshots. Erasing results depend on background complexity - erasing works best with a solid color or a uniformly textured background.
• New Scene Text: Insert specified text into blank areas in images (such as walls, banners, labels). The position of the new text is specified by the user (marked by a rectangular frame or mask), and the model automatically determines the perspective relationship and lighting conditions of the area and generates text effects that match it.
• Multi-language support: Supports scene text editing in Chinese and English. Both languages ​​​​are processed under the same model framework, and Chinese and English text can be mixed and edited in the same image.
• Default font style migration: When editing text, the model automatically extracts the font style of the original text (serif/sans-serif, weight, and tilt angle) and applies it to the target text, without the need for the user to manually specify it.
• Background consistency: Image content outside the editing area is not changed at all - FLUX-Text only operates on the user-specified text area. This is its core difference from general image editing models (these models are prone to "modifying" surrounding areas).

Model and version evolution

Version	Release Date	Key Changes
v0.9 (research version)	2025-05-06	Release arXiv paper and initial reasoning code to verify the feasibility of scene text editing
v1.0 (official version)	2025-07-04	Open Gradio Demo and pre-training weights, support multi-lingual scene text editing

Version records are subject to the official GitHub Release and arXiv paper versions. The iterative direction of FLUX-Text is driven by the research team based on community feedback and technical progress. There is currently no public version roadmap.

Technical advantages

• Diffusion Transformer (DiT) Architecture: FLUX-Text is based on the Diffusion Transformer instead of the traditional U-Net diffusion model. The DiT architecture outperforms similar U-Net models in image generation quality, producing clearer text edges and more natural texture fusion effects. The model uses large-scale image-text pair data + specialized text editing data (including paired images before/after editing) when training.
• Text-aware attention mechanism: The model introduces the attention bias of the text area during the diffusion process - during inference, the model pays more attention to the pixel consistency of the user-specified editing area, while keeping the characteristics of the non-editing area unchanged. This eliminates the problem of "diffusion leakage" (pixels outside the edit area being accidentally modified) at edit boundaries.
• Perspective and geometric adaptation: For oblique text that is not shot from the front (such as perspective text on product packaging), the model automatically detects the geometric transformation parameters (rotation, scaling, perspective matrix) of the text area, and applies the same transformation when generating the target text, so that the edited text is visually consistent with the original perspective angle.
• Conditional control input: Users can precisely specify the editing area through a text mask (binary mask), or use a rectangular box for rough annotation. Masked input allows for arbitrarily shaped editing areas, while rectangular boxes provide a more convenient way to interact.
• Training data strategy: The model uses a hybrid training strategy of synthetic data + real scene data. Synthetic data covers a large number of font, background and perspective changes to ensure the model's versatility; real scene data (real images containing text collected from the Internet) improves the model's generalization ability in real environments.

How to use

How to use	Entrance	Instructions
Gradio Demo	Hugging Face Spaces	Online experience, no local deployment required (there is a queue to wait)
Local inference	GitHub repository	Clone code → Install dependencies → Download weights → Run inference script
Read the paper	arXiv	Understand the technical details and experimental evaluation
Model weights	Hugging Face	Download pre-trained weights for local inference

Typical usage process (local deployment):

git clone https://github.com/AMAP-ML/FluxText
cdFluxText
pip install -r requirements.txt
python run.py --input image.jpg --mask mask.png --target_text "New text content"

Input requirements: an original image containing text + a mask image (or rectangular frame coordinates) identifying the editing area + target text content. Output: edited image file.

Product Pricing

Usage Model	Fees	Description
Gradio Demo (online experience)	Free	Research demonstration use, usage restrictions and waiting queues
Local deployment (self-hosted)	$0 (software) + GPU computing power cost	Software is free and open source, only hardware or cloud GPU costs are borne
Commercial use	Subject to the open source license terms	Confirm the constraints of the license type on the commercial scenario

For individual researchers and AI enthusiasts, Gradio Demo can meet the occasional image text editing needs, but is limited by queuing and concurrency limitations. Local deployment is recommended for batch use in production environments, and the cost of cloud GPU rental is about $0.5-2/hour (depending on the GPU model).

Application scenarios

• Batch modification of posters and advertising images: The marketing team frequently modifies the date, price and event information in the poster before and after major promotions. Using FLUX-Text to replace text in batches is 10-50 times more efficient than manually modifying them one by one in PS. Verification method: Compare the accuracy of batch modification (the text position/size/angle is consistent with the original image) and the manual correction ratio.
• Product image price and parameter update: The e-commerce operation team regularly updates the price, promotion label and specification parameters on the product main image. FLUX-Text can accurately replace price numbers in a specified area, leaving the background and other visual elements unchanged. Verification method: Whether the click-through rate and conversion rate of the edited product image on the platform are affected by editing traces.
• Emoticon and social media graphic creation: Content creators can modify the text bubble content in emoticons and replace the dialogue text in screenshots. Verification method: Whether the editing traces are visually detectable.
• Document and Screenshot Desensitization: Use FLUX-Text to erase sensitive information (name, phone number, email, etc.) before sharing internal document screenshots, which is more natural than manual mosaic. Verification method: Assessment of naturalness of background filling in erased areas.
• Movie and game subtitle replacement: Replace text in video frames (such as replacing Chinese with English in game UI), which needs to be processed frame by frame. Note: The current version is for single image editing, and video processing needs to be performed frame by frame.

Applicable people

Crowd	Adaptation value	Preconditions
AI researcher	Research scene text editing algorithm, which can be used as a baseline comparison	Python + PyTorch environment
E-commerce operation and design team	Batch modification of product images and poster text	Technical students are needed to assist in deploying or using Gradio Demo
Content creator	Emoticon packs and social media graphics and text modification	Gradio Demo is enough for light use
Data desensitization team	Erasure of sensitive information in document screenshots	Local deployment is required to process large amounts of data
Not suitable for the crowd	Reasons
---	---
Ordinary users without technical background	Native usage requires command line operation
Scenes that require video editing	Currently only single images are supported, and videos need to be processed frame by frame
Professional designs that have extremely high requirements for font restoration	The degree of restoration of non-standard fonts may be insufficient

Summary and Outlook

FLUX-Text provides a high-quality open source solution in the highly specialized niche field of scene text editing (STE). Its core value lies in "precisely controlling the text content in the image without affecting other visual elements" - in the past, this either required professional designers to manually process pixel by pixel (a single image took 15-60 minutes), or could only be achieved indirectly through redrawing the entire image but was uncontrollable. FLUX-Text reduces the time required for this operation to seconds while maintaining the professional-grade visual quality of the editing results.

Current limitations: (1) Model inference requires GPU resources (recommended 16GB+ VRAM), which weakens the "zero threshold" advantage of open source - users without suitable hardware can only have limited experience through Gradio Demo; (2) The text editing effect on unconventional fonts (handwriting, artistic typefaces, decorative fonts) is unstable - the model's training data coverage on these fonts may be insufficient; (3) Currently only supports single image processing, and does not support batch processing pipelines or video scene text editing; (4) There may be subtle visual flaws (such as discontinuous strokes and color deviations) in the text output by the model, which may be noticed when viewing enlarged. Purchasing/Adoption Suggestions: The e-commerce and marketing teams recommend first conducting batch testing on a small number of actual pictures (50-100 pictures) to evaluate whether the editing quality meets the business requirements for accuracy and visual naturalness. Research teams can directly download the weights and code to fine-tune on their own data sets. Follow-up directions include: optimization of model inference speed (supporting batch processing), support for more languages, and integration with other image editing capabilities (such as background replacement + text editing combined operations).

Related tools: CrewAI, langchain

Version Info

• FLUX-Text official version ：Open Gradio Demo and pre-trained weights, and support multi-lingual scene text editing. （2025-07-04）
• early research version ：Release the arXiv paper and initial reasoning code to verify the feasibility of scene text editing. （2025-05-06）