Tencent previews Hy Image3.5 for professional visual creation

An image model built around revisions

Tencent's Hunyuan team has introduced Hy Image3.5 preview for professional visual creation. The team's research index dated the item September 21, and the model is now listed in Tencent Cloud's TokenHub documentation. It supports text-driven image creation, generation from reference pictures and successive edits within a conversation.

The meaningful change is the workflow. A designer can describe an initial image and then issue a follow-up instruction with the prior conversation attached. This makes revision part of the same interaction instead of requiring a separate description of the whole composition for every request.

What the API documents

Tencent's image API guide identifies the model as hy-image-v3.5-preview. Requests use a messages structure containing text and optional images. For a later edit, the application carries forward the previous messages and the history returned by the service, then appends the new instruction.

The guide describes support for as many as twenty reference images, with file-size limits, and output with a maximum area equivalent to 4096 by 4096 pixels. Explicit dimensions are required for the highest resolution; automatic sizing uses lower pixel-area tiers. These are interface capabilities documented by Tencent, rather than results from our own image-quality tests.

A preview with practical integration limits

The response includes a generated-image link and a record that can be reused to continue editing. Tencent says the links are temporary, with a twelve-hour validity period. An application that needs a permanent asset library must therefore save its own copy of successful outputs rather than treat the returned URL as durable storage.

The current release is a hosted model. The sources checked for this article do not establish a downloadable weight release. They also do not supply an independent comparison demonstrating that Tencent's model consistently preserves typography, identity or layout better than competitors.

The limits behind the reference-image workflow

Reference images can be supplied through public URLs or Base64 data URLs. The documented ceiling is 20 MB per image and twenty images in the conversation; if the conversation exceeds the image count, older rounds are truncated first. The guide also describes input resizing, with a default reference-image area of about one million pixels to control context consumption. Uploading a large file therefore does not mean every original pixel is passed unchanged to the model.

Output dimensions are governed by both side length and total area. Each side may range from 256 to 8192 pixels, but the area cannot exceed 16,777,216 pixels, equivalent to a 4096-square image. This permits some elongated canvases; it is not permission to generate an 8192-square picture.

For successive edits, the service provides assembled history and a session identifier used to route requests consistently and improve cache reuse. The interface also offers an external-search switch, disabled by default. Applications must explicitly maintain the history and choose whether search is part of the creative workflow.

What professional quality would mean

Multi-turn editing is valuable only when a narrow change leaves unrelated parts of a composition intact. Changing a product's background while preserving its label, proportions and lighting is a different test from generating an attractive picture from scratch.

Tencent's preview provides a concrete interface for that process. Whether it reduces manual cleanup will depend on the kinds of assets and revisions involved, not simply the maximum pixel count. The documentation makes the mechanics inspectable; the quality claims still need evaluation on actual creative jobs.