Tencent Hunyuan has officially launched the HunyuanImage3.0-Instruct model. The model is now available on Tencent's AI assistant "Yuanbao" across all platforms and the official website of Tencent Hunyuan, marking a new breakthrough in Tencent's native multimodal image processing field.

HunyuanImage3.0-Instruct adopts a mainstream Mixture of Experts (MoE) architecture with a total parameter count of 80B (activated parameters around 13B). Unlike traditional filter-based image editing, it is defined as an "intelligent" image editing model. After receiving user prompts and images, the model first deeply understands the content of the image, then autonomously reasons about the areas and steps that need modification, and precisely retains the image details that do not require changes, thereby achieving more logically coherent output results.

image.png

In terms of functionality, the model demonstrates high flexibility. It not only supports basic element addition and removal, style transformation, and old photo restoration, but also has strong multi-image fusion capabilities, enabling the extraction and synthesis of characters or elements from multiple photos. For ordinary users, this means they can quickly create personalized stickers, virtual duets, and even complete professional e-commerce poster designs and character customization directly on Yuanbao.

To refine this model, the Hunyuan team built a large-scale image-to-image dataset covering over 80 specialized tasks. By introducing chain-of-thought training and a self-developed MixGRPO algorithm, the model has significantly improved in instruction response speed and image consistency. Whether in emotional expression or realistic generation, HunyuanImage3.0-Instruct provides a more professional and user-friendly tool for AI image creation.

Key points:

  • 🧠 Autonomous Reasoning Editing: HunyuanImage3.0-Instruct can understand the original image before executing instructions, automatically analyze the areas to modify and maintain consistency in non-edited regions.

  • 🎨 Multi-Scenario Function Coverage: Supports old photo restoration, portrait collage synthesis, and complex text modifications, applicable in various creative fields such as e-commerce posters and game customization.

  • ⚡ Performance and Efficiency Enhancement: Based on an 80B MoE architecture and trained on a dataset of millions of images, the model produces images with stronger emotional tension and significantly faster generation speed compared to previous generations.