Tencent Hunyuan's next-generation image model, Hy Image 3.5 Preview , is now live, focusing on professional-grade AI image generation and editing. According to the release, its win rate in human preference evaluations is about 30% higher than Hy Image 3.0, covering text-to-image, image-to-image, and continuous editing, with a clear shift in goal from "generating a single image" to a complete visual production process.
Not just text-to-image, but also continuous editing
Hy Image3.5 Preview uses Chat/Messages interaction, allowing simultaneous reception of text and reference images, while preserving multiple rounds of context. Users can create images and then continue editing people, scenes, composition, and details via natural language, without having to start over each time.
The latest Tencent Cloud document also shows that the model supports input of up to 20 reference images, with a maximum of 100k tokens for contextual input. Compared to Hy Image 3.0, the number of reference images and the flexibility for complex editing have significantly increased.
The resolution has actually reached 4K
Early releases emphasized up to 2K generation, but Tencent Cloud's current API documentation already supports direct output of 4K images up to 4096×4096 using size parameters; The automatic area settings offer 1K, 1.5K, and 2K options.
The official documentation also emphasizes improved text rendering, realism, and editorial consistency, making it more suitable for scenarios requiring repeated revisions such as posters, UI design, product images, and commercial visuals.
$0.024 per image lowers the cost of image creation
According to this release, the generation price for Hy Image3.5 Preview on Tencent Cloud TokenHub is as low as $0.024 per image, and only the generated results are charged, with no additional charges for uploaded reference images.
Miora and OnSolo also offer a two-week free trial entry. For teams that need a large number of product images, ad creatives, or design iterations, low cost per image plus multiple rounds of editing is more practical than simply pursuing one-time quality.
Generative images are entering a stage of "editability and sustainable iteration." What truly deserves attention in Hy Image3.5 Preview is not just the improved win rate compared to the previous generation, but Tencent's start to integrate high-resolution, multi-reference images, and conversational editing into a unified API, which is closer to what professional visual workflows need.
