Hunyuan-PromptEnhancer/PromptEnhancer

What it solves

PromptEnhancer addresses the gap between simple user prompts and the detailed descriptions required by text-to-image (T2I) models to produce high-quality, accurate images. It prevents the loss of original user intent while transforming vague inputs into structured, descriptive prompts that downstream image generators can better understand.

How it works

The project uses a Chain-of-Thought (CoT) prompt rewriting approach to restructure input prompts. It offers three primary implementation paths:

• Text-to-Image Enhancement: Rewrites text prompts into clearer, structured versions using 7B or 32B parameter models.
• Image-to-Image Editing: Uses a vision-language model (based on Qwen2.5-VL) to refine editing instructions by analyzing the visual context of an input image.
• Quantized Inference: Provides GGUF versions of the 32B model via llama.cpp to allow high-quality prompt enhancement on consumer-grade GPUs with significantly reduced VRAM usage.

Who it’s for

• AI Artists and Designers: Users who want better results from T2I models without manually writing complex prompts.
• Developers: Those building image generation pipelines who need an automated way to refine user input.
• Researchers: Individuals studying prompt engineering and the interaction between vision-language models.

Highlights

• Dual-Mode Support: Handles both initial image generation (T2I) and modification of existing images (Img2Img).
• Intent Preservation: Specifically designed to maintain all key elements like subject, action, style, and layout during the rewriting process.
• Flexible Deployment: Offers a range of model sizes (7B to 32B) and quantization levels (Q4, Q6, Q8) to fit different hardware constraints.
• Robust Parsing: Includes a multi-level fallback mechanism to ensure reliable output generation.