Tencent has introduced Hy-Image 3.5 Preview, a new image-generation model aimed at people who want more control than “type a prompt and hope.” The preview is designed for text-to-image generation, reference-image workflows, and multi-turn editing through a chat-style interface.
That makes it interesting for designers, marketers, product teams, educators, and anyone who needs a visual idea to stay consistent while it changes. It is still a preview, so real-world results and access may vary. But the direction is clear: image creation is moving closer to an editable conversation.
What we verified: Tencent’s official TokenHub documentation lists the model as hy-image-v3.5-preview and describes its supported inputs, image limits, editing flow, sizing controls, and output behavior. See the official Tencent Cloud guide.
What is Hy-Image 3.5 Preview?
Hy-Image 3.5 Preview is part of Tencent’s Hunyuan image-generation family. In simple terms, it turns a written idea—and optionally one or more reference images—into a new image. The important difference is the interaction model: Tencent documents a Chat/Messages-style API where a request can carry conversation history for follow-up edits.
You can start with a brief, ask for a different composition, remove an object, change the mood, or keep a subject while moving it into a new scene. The system is intended to remember the earlier visual direction during that session.
The features that matter most
1. Text-to-image generation
The familiar workflow is still here: describe the subject, style, lighting, composition, and intended use. The service can return a finished image synchronously, which means an application can wait for the result instead of building a separate job-and-polling flow.
2. Reference images and image editing
Hy-Image 3.5 Preview accepts text plus reference images. Tencent’s guide says it supports up to 20 reference images per request, with each image up to 20 MB, using a public URL or base64 data. This is useful when a prompt alone is not enough to describe a product, person, room, outfit, logo direction, or visual style.
For a small business, that could mean keeping the same product shape while trying several backgrounds. For a creator, it could mean using a character sheet or a mood board as visual guidance.
3. Multi-turn editing
This is the feature that changes the feel of the tool. Instead of restarting every time, you can treat the image as a draft and continue the conversation: “keep the cup and mountains, make the light warmer,” or “use the same character, but show a wider view.”
Tencent’s documentation explains that previous messages and generated-image history can be sent back for the next turn. For developers, this is a practical way to build an editor that feels more like a chat than a collection of disconnected buttons.
4. Up to 4K output
The preview supports custom dimensions and can reach up to 4096 × 4096 pixels when the requested area stays within the documented limit. If you do not provide an exact size, the service can choose from area settings such as 1K, 1.5K, or 2K. For a precise 4K result, Tencent recommends setting the size directly.
Higher resolution does not automatically mean a better picture. It is most useful when the image will be cropped, printed, used in a large banner, or placed inside a product design where small details matter.
5. Optional search enhancement
The documented API includes a switch for external search enhancement. That suggests a workflow where a model can use current information to improve a visual brief, although the feature is listed as off by default. Teams should treat this as a deliberate choice: search can add useful context, but it also creates a need for clearer privacy, source, and review rules.
What is different from Hy-Image 3.0?
Tencent’s documentation presents Hy-Image 3.0 as a simpler image-generation API with reference-image support, while Hy-Image 3.5 Preview adds a chat-style messages protocol, multi-turn context, more reference images, and a larger output ceiling.
The practical difference is not just a bigger number. Hy-Image 3.5 Preview is built for a longer creative process: brief, draft, correction, refinement, and delivery. That makes it a better fit for visual work where the first result is only the beginning.
Who should try it?
- Designers: explore layouts, art directions, and visual variations before polishing a final asset.
- Marketing teams: create a family of campaign images while keeping a product or visual idea recognizable.
- Content creators: make thumbnails, article illustrations, social visuals, and story concepts.
- Developers: build image editors and creative tools around an OpenAI-style messages workflow.
- Small businesses: turn a simple product description and a few reference photos into more useful visual options.
Important limits and questions
Because this is a preview, it should be tested before it is used for a high-volume production pipeline. Generated image URLs are temporary according to Tencent’s documentation, so applications need to download and store results rather than relying on the returned link forever.
Teams should also check commercial-use terms, regional availability, content-safety behavior, latency, and the cost of their chosen service route. A strong demo is not the same as a dependable production workflow.
Hy-Image 3.5 Preview is most interesting when you stop judging it as a prompt box and start judging it as a visual editing partner.
Our early take
Hy-Image 3.5 Preview looks like a meaningful step toward more conversational image tools. Its combination of reference images, multi-turn editing, explicit size control, and 4K output targets the messy middle between a rough idea and a finished visual.
For everyday users, the promise is simple: fewer restarts and more useful revisions. For developers, the appeal is a structured workflow that can fit into an image product. The preview label still matters, but this is a release worth watching if your work depends on images.
Bottom line: Hy-Image 3.5 Preview is not just about generating more pictures. It is about keeping the idea alive while the picture gets better.
Sources: Tencent Cloud TokenHub Hy image guide and the public Hy Image 3.5 product overview. Details can change while the model is in preview.
