PromptEnhancer - Tencent Mixed Meta Open Source AI Prompt Word Enhancement Tool

What is PromptEnhancer

PromptEnhancer is an open source prompt word enhancement tool from Tencent's Mixed Meta team to improve the generation of text-to-image (Text-to-Image, T2I) models. Through Chain-of-Thought (CoT) approach to reconstruct the user input of simple prompt words, to generate richer and clearer prompt words, so that the T2I model more accurately understand the user's intent, to generate more compliant images.PromptEnhancer equipped with a reward model called AlignEvaluator, the model PromptEnhancer is equipped with a reward model called AlignEvaluator, which evaluates generated (image, cue) pairs based on 24 fine-grained keypoints, and outputs scalar reward signals to guide the rewrite of the model for optimization. It can be used as a generalized cue word enhancement framework to improve model performance without modifying the weights of pre-trained T2I models. Multiple output parsing methods and configurable inference parameters are supported to meet different user needs.

Features of PromptEnhancer

• Cue word optimization: It can reconstruct the simple cue words entered by the user into richer and clearer cue words, improve the text-to-image model's understanding of the user's intent, and generate more compliant images.
• Chained reasoning rewritten: The rewriting of cues using Chain-of-Thought (CoT) to make the generated cues more logical and structured.
• Semantic Alignment Evaluation: Equipped with the AlignEvaluator reward model, which evaluates generated (image, cue) pairs based on 24 fine-grained keypoints and outputs a scalar reward signal that guides the rewrite model optimization.
• universal adaptationThe T2I model can be used as a generalized cue word enhancement framework to adapt to a variety of pre-trained models, such as hybrid, Stable Diffusion, etc., without modifying the weights of the pre-trained T2I model, which reduces the cost of optimization.
• Multi-language support: Supports bi-directional conversion between English and Chinese to avoid ambiguity of expression due to language differences and to enhance the effect of cross-language generation.
• interpretabilityThe CoT chain of thought and 24-dimensional evaluation make the cue optimization process more transparent, and developers can clearly locate the blind spots of model understanding.
• Configurable parameters: Users can adjust parameters such as temperature, top_p, and maximum number of newly generated tags as needed, balancing the certainty and diversity of the generated results.
• ecological complementation: The team released a high-quality human preference benchmark containing a large amount of labeled data for complex scenarios, which provides an important reference for subsequent cue optimization studies.

PromptEnhancer's Core Benefits

• Significantly improves image generation: By optimizing the cue words, the consistency of the generated images with the textual descriptions is dramatically improved, especially in complex scenes and detailed representations.
• No need to modify model weights: As a plug-and-play module, it does not require weight modification of the pre-trained T2I model to achieve performance improvement and reduce the optimization cost.
• Multi-language conversion supportIt has the ability of bi-directional conversion between Chinese and English, which effectively avoids the ambiguity of expression caused by language differences and expands its application scope in different language environments.
• Equipped with professional assessment models: Built-in AlignEvaluator reward model that evaluates generated results at 24 fine-grained key points to ensure accuracy and effectiveness of optimization direction.
• Enhanced interpretabilityThe CoT thought chain and multi-dimensional evaluation mechanism make the cue optimization process more transparent, making it easier for developers to locate and solve the blind spots in model understanding.
• Provision of high-quality baseline data: The team released high-quality human preference benchmark data for complex scenarios, which provides an important reference and support for subsequent research and optimization.

What is PromptEnhancer's official website?

• Project website:: https://hunyuan-promptenhancer.github.io/
• Github repository:: https://github.com/Hunyuan-PromptEnhancer/PromptEnhancer
• HuggingFace Model Library:: https://huggingface.co/tencent/HunyuanImage-2.1/tree/main/reprompt
• arXiv Technical Paper:: https://www.arxiv.org/pdf/2509.04545

People for whom PromptEnhancer is intended

• content creator: Artists, designers, ad creators, etc. who need to create visual content by generating images from text, use PromptEnhancer to optimize cue words and generate images that better meet creative needs.
• AI developers: Professionals working to improve the performance of text-to-image models can use PromptEnhancer as a tool to optimize cue words and improve model generation without modifying model weights.
• research workerScholars working at the intersection of natural language processing and computer vision can use PromptEnhancer to explore the impact of cue word optimization on model performance and promote the development of related technologies.
• creative workerPromptEnhancer helps writers, screenwriters, etc., who need images to assist them in conceptualizing their ideas, to transform their written ideas into visual images with greater precision and to inspire more creative ideas.
• Students and educatorsPromptEnhancer can be used to optimize prompt words and generate images to assist teaching or learning, and to improve the understanding and expression of complex concepts.