MiniMax H3-Context-IR returns a prompt, not a video: Sume view

H3-Context-IR is a MiniMax endpoint, POST /v2/h3_context_ir, that reads your text, images, video and audio and returns an enhanced video prompt without creating a video. Sume's docs list no equivalent standalone prompt-enhancer endpoint, so on Sume you write the prompt yourself or ask an agent to write it, then submit minimax-h3 with the finished text.

Everything about the MiniMax endpoint below is from its page, Create H3-Context-IR Task, read on 2026-10-02. Sume's side is from Video generation and Video Router.

What does H3-Context-IR actually do?

According to the page, it interprets multimodal context across text, images, audio and video and generates an enhanced video prompt, and it does not create a video generation task. The request takes model (only MiniMax-H3), a content array, a required duration of 4 to 15 seconds, an optional ratio, and an optional callback_url. It returns a task_id with task type h3_context_ir; you fetch the finished prompt from content.prompt through the query-task endpoint.

It uses the same media limits as video generation: images up to 30 MB, videos up to 50 MB and 3 clips, audio up to 15 MB and 3 clips, and 64 MB for the whole request. The text item is limited to 7,000 characters.

Does Sume expose anything like it?

In the pages I read, no. Sume's /v1/videos takes prompt, duration, resolution, aspect_ratio, frame_images, input_references, generate_audio and callback_url. provider.options is rejected with 400 unsupported_parameter because the passthrough list is empty in v1 for every model, and seed and size are rejected as well. There is no endpoint that returns only a prompt.

That also means MiniMax's extra generation options, such as prompt expansion mode, are not reachable through Sume. If you need the enhancer specifically, you would call MiniMax directly; Sume will not proxy that call.

What is the Sume-side workflow instead?

Do the context reading yourself. Describe each reference by role in the prompt, send the references as input_references or Video Router reference_*_urls, and keep the prompt under MiniMax's 7,000 characters. Sume's guide says reference-to-video treats images as visual guidance, not exact frames, and that frame_images takes precedence if you send both. Audio and video references are honored by minimax-h3 and minimax-h3-max.

A concrete pattern: ask your own assistant to draft the prompt from your assets, review it, then submit once. You pay for one video, not for a draft video.

import os, requests

prompt = open("reviewed-prompt.txt").read().strip()
assert 0 < len(prompt) <= 7000
r = requests.post(
    "https://api.sume.com/v1/video-router/generate",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}",
             "Idempotency-Key": "h3-ctx-001"},
    json={"model": "minimax-h3", "prompt": prompt, "duration": 8,
          "resolution": "768p",
          "reference_image_urls": ["https://example.com/product.png"],
          "mode": "async"},
)
print(r.status_code, r.json())

When is skipping the enhancer a loss?

If your inputs are a mix of long audio and video and your prompts come out vague, a vendor-side reader may catch details you skip. The honest trade is control and a single bill on Sume against an extra vendor hop elsewhere. For the writing side, see the H3 negative-direction prompt guide, and for what the switch does and does not do on Sume, the prompt expansion post.