Tencent Hunyuan
Tencent Hunyuan's PromptEnhancer is an open research project and prompt-rewriting utility for text-to-image generation and image-to-image editing. It converts concise or ambiguous user instructions into clearer, structured prompts while preserving intent, helping downstream generative-image models render attributes, relationships, layouts, style and text more faithfully.
At a Glance
- AI-image researchers and developers
- Creators using text-to-image and image-editing systems
- Open-source and self-hosting communities
- Teams building image-generation applications that need improved prompt adherence
AI Tools by Tencent Hunyuan
(1)
PromptEnhancer
Text to Image Prompt Rewriter
Discussions
No discussions yet
Be the first to start a discussion about Tencent Hunyuan
Latest News
Released the T2I-Keypoints-Eval evaluation script.
PromptEnhancer presented at CVPR 2026 as a poster on taming prompt rewriters with fine-grained reward.
Released the PromptEnhancer-32B Gradio demo.
Released the PromptEnhancer image-to-image editing model.
Products & Services
A 7B prompt-rewriting model intended as the balance-of-quality-and-efficiency option. It supports Chinese and English text-to-image prompt enhancement and can be run locally through the project's Python inference code.
A larger multimodal language model for higher-quality text-to-image prompt enhancement and chain-of-thought rewriting, distributed through Hugging Face with local inference support.
A vision-language prompt-rewriting model that uses both an input image and an editing instruction to produce clearer, structured and logically consistent image-editing prompts while preserving intent.
A human-aligned evaluation dataset for text-to-image prompt following, covering fine-grained semantic and compositional capabilities in Chinese and English; the project later released an evaluation script.
Market Position
PromptEnhancer is positioned as a universal, model-agnostic prompt-rewriting layer rather than a new image generator: it can improve a pretrained T2I model without retraining or changing generator weights. Its differentiation is chain-of-thought restructuring plus reward optimization with a 24-keypoint AlignEvaluator, compared with earlier model-specific or implicitly CLIP-rewarded rewriters and general-purpose prompt-enhancement tools. The project is open-source/self-hostable and offers both T2I and image-editing variants.
Founding Story
The project was started to address the prompt-alignment gap in text-to-image systems: models frequently misinterpret attribute binding, negation, object counts, spatial relationships and other compositional instructions. The initial vision was a universal, model-agnostic rewriter that could improve any pretrained T2I generator without changing its weights, using chain-of-thought rewriting and a fine-grained AlignEvaluator reward model.
Target Markets
- AI-image researchers and developers
- Creators using text-to-image and image-editing systems
- Open-source and self-hosting communities
- Teams building image-generation applications that need improved prompt adherence
- Improving prompts for text-to-image generation
- Generating photorealistic, digital-art, abstract-geometry and other styled images from concise instructions
- Complex compositional scenes involving attributes, object counts, actions, spatial relationships or negation
- Posters, slogans, in-image text and multi-panel comics
- Image-to-image editing where the instruction must be grounded in visual context
- Research benchmarking of text-to-image prompt following