HunyuanImage
A look at the open HunyuanImage 3.0 and its Instruct variant: how an 80B MoE text-to-image model follows long prompts, and where image-to-image editing and multi-photo fusion still have limits.
HunyuanImage 3.0 opens up Tencent's image generation models with an 80B MoE architecture that trains text and images together, activating roughly 13B parameters per token at inference. It follows long prompts and renders readable text inside posters and interface mockups, while the Instruct variant turns to image-to-image edits such as adding elements, restyling scenes and merging several photos.
