On January 26, Tencent Hunyuan officially launched its latest **Hunyuan Image 3.0 Image-to-Image (HunyuanImage3.0-Instruct)** model. This model fills an important gap in Hunyuan's image editing and multi-image fusion areas, and is now available on Tencent Yuanbao across all platforms and the official website.

QQ20260126-150904.png

80B Parameter Giant: Intelligent and Instruction-Following Image Editing

Hunyuan Image 3.0 Image-to-Image model has a total parameter count of 80B (about 13B activated parameters), using an advanced Mixture of Experts (MoE) architecture. Unlike traditional photo editing tools, this model has outstanding semantic understanding and reasoning capabilities: upon receiving an image and prompt, it first analyzes the image content and breaks down the editing steps, accurately identifying the areas to be modified and retained.

Based on a massive amount of image-to-image data and the self-developed MixGRPO algorithm, the model has made significant improvements in maintaining consistency in non-edited areas, instruction response speed, and realistic performance.

QQ20260126-150913.png

QQ20260126-150920.png

Full Scene Coverage: From Old Photo Restoration to Multi-Photo Collages

This model supports over 80 sub-tasks, offering users diverse creative capabilities:

  • Image Refinement: Supports adding, removing, or modifying lighting, skin, and background, even changing text in the image without disrupting the layout.

  • Style Transformation: Achieves transforming real people into cartoons, converting to pixel styles, or restoring old photos.

  • Deep Integration: Supports extracting characters or elements from multiple photos to quickly create new collages or creative images.

Currently, users can experience AI-made emoticons, virtual duets, and e-commerce poster design features through Tencent Yuanbao. This marks the official entry of large models into the efficient "intention-driven" era of image editing.