Tencent drops HunyuanImage 3.0 with built-in chain-of-thought image editing

HunyuanImage 3.0-Instruct is Tencent's newly open-sourced image-to-image model. Simply put, it lets you edit pictures or blend multiple images together using natural language instructions, without destroying the parts you wanted to keep.

What you can do with it

Most text-to-image models struggle with precise image editing. HunyuanImage 3.0-Instruct tackles this by specializing in instruction-following and visual reasoning.

For single-image edits, you can add, remove, or swap specific objects, like turning a croissant on a plate into lasagna. It handles style transfers, turns photos into painted portraits, colorizes black-and-white archives, and can even alter complex scene dynamics like camera perspective (say, switching to a top-down view) or lighting conditions.

It also handles multi-image fusion. You can feed it portraits of different people or isolated objects from separate shots and combine them into a coherent single scene or group photo.

How the reasoning engine works

Standard diffusion models usually guess what you mean right away, often ruining the untouched background in the process. HunyuanImage 3.0-Instruct behaves more like a reasoning model with built-in chain-of-thought (CoT).

When you give it an image and a prompt, it pauses to think. It analyzes the image content, deduces exactly which regions need edits, outlines the step-by-step modifications, and locks down the areas that must stay identical before generating pixels.

According to IT Home, the model uses a Mixture-of-Experts (MoE) architecture with 8 billion total parameters and roughly 1.3 billion active parameters, though some community materials cite total parameters up to 80 billion. Tencent trained it on tens of millions of image-to-image pairs covering over 80 tasks, using its proprietary MixGRPO algorithm with reward models during post-training to align outputs with user preferences.

Hardware specs and where to run it

While the weights are fully open-sourced, self-hosting is definitely not for the faint of heart. The entire checkpoint weighs around 169 GB, demanding massive enterprise-grade VRAM that leaves consumer graphics cards out of the question.

If you have the cluster capacity, it serves as a powerful self-hosted editing backend. If not, the easiest way to try it right now is through Tencent Yuanbao across mobile and desktop, or via the official Tencent Hunyuan web interface, where features like photo restoration, sticker making, and e-commerce poster design are already live.

Why it matters

Precise open-source image editing has long lagged behind text-to-image generation. Most existing open tools treat prompt editing like a blunt repaint, failing to preserve subject consistency.

Tencent's move to insert a chain-of-thought planning layer directly into image editing solves a major headache in AI workflows. While the massive 169 GB footprint locks out indie hobbyists on consumer hardware, enterprise teams and studios finally get an open model capable of surgical image manipulation without relying on closed commercial APIs.