Tencent adds conversational editing to Hy Image 3.5 Preview

Tencent has released Hy Image 3.5 Preview, an image-generation API that lets developers refine pictures through a continuing conversation. The service costs $0.024 per generated image and supports text prompts, reference images, and multi-turn editing. Ryan Merket reports for RuntimeWire that Tencent is positioning the model for creative software where users make several revisions to the same asset.

Instead of treating every image request as a separate task, Hy Image 3.5 accepts an ordered history of messages. Developers can send text instructions, reference images, and previous editing turns together. The API then returns an assembled history that can be used in the next request. This should reduce the work required to preserve context during a sequence of changes.

Tencent says the model supports up to 20 reference images per request. Each image can be as large as 20 MB. The documented input context reaches 100,000 tokens, which gives applications room to retain detailed instructions and prior exchanges.

More than the advertised 2K limit

Tencent markets the service around output of up to 2K resolution. Its documentation, however, describes a separate size setting that can create images up to 4,096 by 4,096 pixels. Individual dimensions may reach 8,192 pixels, as long as the total area stays below 16,777,216 pixels.

The API responds synchronously, meaning applications receive the result without polling a separate job. Generated image links expire after 12 hours, so teams need to store final files themselves. Tencent also offers fixed seeds for reproducible outputs, custom watermark text, and an option for external-search enhancement.

Tencent claims Hy Image 3.5 scores 30% higher than Hy Image 3.0 in its own human evaluations. That figure is not an independent benchmark. The lower price, which Tencent lists as 25% below Hy Image 3.0’s $0.032 reference price, may nevertheless make practical testing easier for teams assessing image consistency, text rendering, and reference-image preservation.