Hy Image 3.5 API guide: calling hy-image-v3.5-preview

Quick facts

hyimage35.pro does not offer a developer API for Hy Image 3.5 and does not resell Tencent Cloud access. To call the model from code you need your own TokenHub key. The workspace on this site is for generating in the browser.

The request body

Unlike the older Hy-Image-3.0 endpoint, which takes a single prompt string, the 3.5 endpoint takes a messages array. Each message has a role and a content list whose parts are either {"type": "text"} or {"type": "image_url"}. An image part accepts an HTTP(S) URL or a Base64 data URI.

curl -X POST 'https://tokenhub-intl.tencentcloudmaas.com/v1/wand/hunyuan-image/v35-generation' \
  -H 'Authorization: Bearer YOUR_API_KEY' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "hy-image-v3.5-preview",
    "generate_max_pixels": 4194304,
    "messages": [
      {
        "role": "user",
        "content": [
          { "type": "text", "text": "Keep the carafe and the light. Replace the ledge with dark slate." },
          { "type": "image_url", "image_url": { "url": "https://example.com/carafe.png" } }
        ]
      }
    ]
  }'

The response is a single JSON object rather than a stream, even though the image sits under a delta key. Alongside the image it returns request_id, a tokenhub_usage.total_tokens figure that the guide describes as the number used for billing reconciliation, and assembled_history.

Output size: 1K, 1.5K, 2K and explicit 4K

There are two ways to control size, and they do not combine.

• generate_max_pixels picks a target area and lets the model choose the shape. It accepts three values: 1048576 (1K), 2359296 (1.5K, the default) and 4194304 (2K). It only takes effect when size is not sent.
• size sets exact dimensions as WIDTHxHEIGHT. Each side may be 256 to 8192 pixels and the total area may not exceed 16,777,216 pixels, which is where the 4096x4096 ceiling comes from.

So "up to 2K" and "up to 4K" are both accurate descriptions of the same API. The automatic tiers stop at 2K, which is also what Tencent's launch post advertises. Output at 4096x4096 is available when you ask for it explicitly with size.

Reference images

The API guide allows up to 20 reference images per request, in PNG or JPEG, at up to 20 MB each. resize_max_pixels sets the maximum area a reference is resized to before the model reads it and defaults to 1,048,576 pixels. The prompt side accepts up to 100k tokens.

Tencent's launch materials and the ComfyUI partner node describe five reference images, not twenty. Treat five as the limit inside products and integrations, and twenty as the documented API ceiling.

Multi-round editing with assembled_history

Multi-round editing is the feature the messages format exists for. Each response includes assembled_history: a ready-made sequence of messages recording what the model did in that round, including the image it produced. The guide's instruction is to append that array unchanged to your messages and then add the next user turn.

• Round one: send the brief, with references if you have them. Keep the returned assembled_history.
• Round two: send the original messages, then assembled_history, then a new user message that states one change and what must stay.
• Repeat. The model reads the earlier image and instructions from the history instead of starting over.

Two details matter. Generated images fed back through the history are marked is_generate: true on a role: "tool" message; you should not set that flag on your own uploads. And the optional session string keeps related requests on consistent routing, which the guide says improves cache hits across rounds.

Other parameters and errors

Documented error statuses are 400 for a malformed request, 401 for a failed key, 422 when content moderation blocks the request, 429 when you exceed concurrency, and 500 for server errors. The guide recommends backing off and retrying on 429. One third-party hands-on from GMI Cloud reported that identical seeds did not always return identical images during the preview, so verify reproducibility before you depend on it.

Pricing and where else the API is offered

Tencent's launch post prices Hy Image 3.5 Preview at $0.024 per image on the Tencent Cloud API, and coverage of the release states that only output images are charged, so reference images do not add cost. The API guide itself does not carry a price table; check the TokenHub console for the rate that applies to your account and region.

Outside Tencent Cloud, GMI Cloud lists the model in its inference engine under the same id and at the same $0.024 per image, and ComfyUI exposes it through partner nodes billed through a Comfy account. When this page was reviewed we had not found it on WaveSpeed, fal, Replicate or OpenRouter.

Use Hy Image 3.5 in ComfyUI Every way to try the model