Tencent released Hy Image 3.5 Preview on 22 September and said an internal test involving hundreds of its designers found its image output on par with ByteDance’s Seedream 5.0 Pro, Bloomberg reported. Tencent also rated its model slightly above Google’s Nano Banana Pro and Alibaba’s Qwen-Image-3.0 Pro, though it did not quantify those quality comparisons.
Key points
- Tencent says a blind evaluation by hundreds of its designers put Hy Image 3.5 Preview level with ByteDance’s Seedream 5.0 Pro.
- The company gave no numerical measure for its comparisons with rival models.
- The model accepts text instructions and up to five reference images, and is being integrated into Tencent products.
Tencent’s blind test and rival models
The comparison came from Tencent’s own blind evaluation, in which hundreds of in-house designers assessed image output, KrASIA reported. Tencent said the exercise placed Hy Image 3.5 Preview alongside Seedream 5.0 Pro and slightly ahead of Nano Banana Pro and Qwen-Image-3.0 Pro. That gives the launch a specific set of rivals, but Bloomberg reported that Tencent supplied no numerical measure for those quality claims.
Tencent separately reported a 30% improvement over Hy Image 3.0 in its own human evaluation, according to KrASIA. That figure concerns Tencent’s previous model, rather than the comparison with ByteDance, Google and Alibaba. For a company putting image generation into products used for editing and design, the internal assessments offer a way to judge whether an update is worth deploying across different tasks.
The release took place on the day Alibaba opened its AI conference. Tencent shares gained more than 7% in Hong Kong, Bloomberg reported. The company has also been changing how it develops models: Yao Shunyu, a former OpenAI researcher, overhauled its AI infrastructure and benchmarking standards and now serves as chief AI scientist. He has publicly argued against training solely for leaderboard scores, favouring integration into products and work on practical problems.
Five reference images and Tencent’s products
Hy Image 3.5 Preview can generate an image from text or edit an existing image, and it accepts up to five reference images, KrASIA reported. A reference image gives the model something visual to work from when an instruction alone would leave too much open to interpretation. Tencent says the update concentrates on readable text and layout, realistic or stylised output, and consistency when an image is edited — qualities that matter when a picture must retain selected details through a change.
Making a poster with readable words could be easier if the model follows the requested layout and renders its text clearly. Revising an image could also require less reworking if selected details remain consistent, though Tencent’s comparisons with rival models give no numerical measure of those qualities.
Tencent is integrating the model into Yuanbao, its chatbot, as well as film-editing and design tools. Other integrations include content creation, advertising, workplace and cloud products. Tencent calls its approach co-design: product teams use the model on particular tasks and work with the model team on the problems they encounter. The same products then become places where the revised model can be used.
Tencent Cloud offers Hy Image 3.5 at USD 0.024 per 2K output image, with charges applying to images produced, KrASIA reported. That price attaches a direct cost to each generated image, while the supported reference images and editing requests give developers ways to fit generation into work that starts with existing visual material.
Yuanbao requests during a limited rollout
Tencent reported more than 50% growth in overall image-generation requests during a limited Yuanbao rollout of Hy Image 3.5, according to KrASIA. It said portrait-retouching requests doubled and requests for work tasks such as posters and logos rose 80%. Those are figures for requests in one product during that rollout, a different measure from designers’ judgements of image quality.
Other Tencent teams have been testing narrower uses. The ima knowledge assistant integrated the model a month before its release and worked with the Hy team on illustrations, posters, infographics and editing. Tencent said use of ima’s image-generation feature subsequently tripled. WorkRally tested the model with studios making animated short dramas, and its product team said it could produce batches of props, characters and scene assets while meeting most of its requirements for following instructions and keeping characters consistent.
Tencent said its advertising team saw a 7% reduction, against Hy Image 3.0, in image-to-image problems including garbled text and abnormal redrawing.
