A designer opens a tool, types "mid-autumn seafood safety infographic, 3:4, warm journal style," and expects a printable poster — not a pretty but useless picture with garbled text. For most image models, that last mile has been the hard part. Tencent's newest release is an explicit attempt to close it.
Key takeaways
- Tencent Hunyuan released Hy Image 3.5 preview on September 22, 2026, a professional-grade text-to-image and image-to-image model.
- In blind tests with hundreds of professional designers, the model's win rate was about 30% higher than its Hy Image 3.0 predecessor.
- It supports up to 5 reference images, multi-round contextual editing, multiple aspect ratios, and output up to 2K resolution.
- The model is wired into Tencent's own productivity stack — Yuanbao, the WorkRally film platform, OnSolo, Miora, WorkBuddy, and ima — rather than standing alone.
The text-rendering problem
Image generation models are good at faces and landscapes and notoriously bad at words. Put "SALE 50% OFF" on a billboard in a generated ad and you often get confident nonsense. Hy Image 3.5 preview leans hard into fixing this: Tencent says it improves rendering of both Chinese and English text, especially small fonts, and strengthens long-text, complex layouts, and mixed image-text composition. For posters, knowledge cards, infographics, and UI design, that is the feature that decides whether a output is usable or decorative.
Built for iteration, not one-shots
A second theme is editability. The model accepts up to 5 reference images and can carry context across multiple editing rounds, so a creator can refine a character or product without starting over. It outputs multiple aspect ratios and up to 2K resolution, which matters less for social thumbnails and more for print and film production. Tencent is clearly pitching this to people who ship deliverables, not people who generate for fun.
Where it actually lives
Hy Image 3.5 preview is not a standalone web toy. Tencent has routed it through its own productivity products. The most telling integration is WorkRally, Tencent Video's industrial-grade AI platform for premium content, which got first access and opened a two-week free trial to creators on September 22. WorkRally covers script analysis, storyboard generation, production, and asset management; plugging Hy Image 3.5 into it means the model feeds directly into storyboards, film materials, and character or scene design. It also connects to Yuanbao, OnSolo, Miora, WorkBuddy, and ima, spreading the model across chat, creation, and document workflows.
What "professional" really means here
The term gets thrown around, but in Tencent's framing it has specific content. A professional image model is judged less on a single striking output and more on repeatability: can it keep a product's color identical across fifty frames, render a price tag without typo-ing it, and re-edit one region without disturbing the rest? Those are the failures that stop a tool from leaving the prototype phase. Hy Image 3.5 preview's emphasis on up to 5 reference images, contextual multi-round editing, and accurate text and layout is a direct answer to that bar. For e-commerce, where thousands of SKU images need a consistent look, and for film, where storyboards must match a final aesthetic, the value is in the boring reliability, not the spectacle.
The Chinese image-model race
Tencent is not alone. Chinese internet giants and dedicated AI labs have all shipped image foundations in the past two years, each aiming at the same creative-professional wedge. The interesting shift is the move away from "who makes the prettiest picture" toward "whose model fits a real workflow." Tencent's answer is to embed Hy Image 3.5 preview inside products people already use for work, rather than asking them to adopt a new standalone app. Whether that distribution advantage outweighs rivals' model quality will be decided by designers, not benchmark charts.
Why a "preview" still matters
The "preview" label is worth taking seriously. Tencent is signaling that this is an early, evolving build rather than a frozen product — and that the real test is professional feedback, which is exactly the blind-test loop it cites. The 30% figure comes from the company's own designer evaluations, so it is a directional claim, not an independent benchmark. But the strategy is coherent: Chinese labs are racing to move image generation from consumer novelty into paid creative workflows where accuracy and consistency justify a subscription.
Honest limitations
The 30% win-rate claim is Tencent's internal blind test; we found no independent third-party benchmark published at the time of writing, so it should be read as the vendor's assertion. Because this is a "preview," capabilities and availability may change, and the two-week WorkRally free trial was a time-limited launch offer rather than a permanent price. We have not tested output quality firsthand. Our coverage is confined to officially disclosed specs and Chinese tech-press reporting; behavior for non-Chinese design contexts may differ.
What readers can do now
- Designers should try Hy Image 3.5 preview through WorkRally or Yuanbao and stress-test it on the task image models usually fail: real text, precise layout, and brand-consistent edits.
- Teams producing film or ad materials can use the multi-round editing to lock a character or product look once and reuse it across storyboards.
- Anyone evaluating Chinese image models should compare this against rivals on text-heavy and layout-heavy prompts specifically, where the gap tends to show.