Media API reference inputs: preflight images and MiniMax H3 assets

A valid task JSON body does not prove its reference assets can be read or used. Before a Nano Banana, GPT Image or MiniMax H3 request, validate the asset and the role it plays in the request. This preflight workflow complements submission and polling; it does not add an upload endpoint or change platform restrictions.

Separate four kinds of validation

Layer	Question	What passing does not prove
Authorization	May this asset be used for this task?	That its URL is reachable
Transport	Does the URL serve the intended bytes?	That the media is decodable
File	Can the image/video/audio be opened with the expected type?	That the model accepts its dimensions or duration
Contract	Does this input mode and combination satisfy the current schema?	That the generated result will match the brief

Open your authorized asset in a local media viewer and inspect its actual format, dimensions and, where relevant, duration and audio tracks. Do not infer the type from a URL suffix. A response containing an HTML login page is not an image even if the URL ends in .png.

Keep a manifest for the exact asset version

Store an application asset ID, version or checksum, intended role, measured file properties and URL expiry. Keep signed URLs and credentials out of public logs. The manifest is application metadata, not additional fields accepted by input.

Browser access can depend on cookies or a local network. Check that the generation service can fetch the intended asset using its supported input method; a successful browser preview is not sufficient. Signed URLs must remain valid for the required processing window, including possible queueing. There is no universal safe expiry duration: use the documented route behavior and an operational margin appropriate to your application.

Do not forward your platform API key to an asset host. If an asset location is rejected by network policy, use a supported authorized hosting path rather than bypassing the policy. Keep private material on access-controlled storage with the narrow access needed for processing.

Validate the selected H3 mode as a whole

The following is an illustrative 768P H3 start/end-frame task. Replace both placeholders with authorized reachable images and validate the complete live schema before a paid submission.

{
  "model": "minimax-h3",
  "input": {
    "prompt": "A single continuous camera move from the opening desk view to the closing close-up. Preserve the same lamp shape and cream material. No extra objects or cuts.",
    "frames_images": [
      "https://assets.example.com/lamp-start.png",
      "https://assets.example.com/lamp-end.png"
    ],
    "seconds": 5,
    "resolution": "768P",
    "aspect_ratio": "16:9"
  }
}

frames_images order is meaningful: start then end. Do not add H3 images, videos or audios to this frame-mode request. For general references, use the separate supported mode and verify its combination rules. Nano Banana and GPT Image use images for their reference-image task inputs; do not copy the H3 field name into those requests. See H3 mode selection and image specifications.

Diagnose without duplicate submission

If a task ID was returned, retain it and inspect the original task. A later asset error is different from submission never reaching the server. Record the task ID, asset version, expiry window and sanitized error. Check source availability and format before deciding on a new attempt. Do not assume an error means no charge; reconcile the actual task record.

If generation completes but the wrong subject appears, review the chosen reference version, mode and prompt. That is a creative or input-selection problem, not evidence that the transport failed. Correcting the brief and generating again is a new budgeted attempt.

Reuse checked assets deliberately

For a batch, freeze an approved asset version instead of silently replacing a file behind the same URL. Recheck expiry before later submissions. Keep accepted outputs linked to the original reference version and budget record; this makes regressions and creative changes distinguishable.

Platform field checks dated 2026-11-03 follow repository schemas and the H3, Nano Banana 2 and GPT Image 2.5 Flare guides. MiniMax’s native API is a conceptual reference; native upload mechanisms and limits are not automatically platform features. No universal file-size or URL-expiry allowance is asserted here.