MiniMax H3 Generate Video (Ace Data Cloud)

This node renders nothing on your machine. It packages your prompt into an HTTPS request, sends it to Ace Data Cloud, and hands you back a task ID you then have to go and collect. No GPU, no checkpoint, no 42 GB download - and no refund if you queue it five times because you got impatient.

H3 makes that trade more tempting than usual. MiniMax H3 is a 33B omni-modal video model with native stereo audio, 4–15 second clips at up to 2K, and it took #1 in Video Editing (With Audio) on Artificial Analysis at launch. The open weights dropped in August 2026 - but the Community License's Applicable Territory excludes the US, the EU, the UK and South Korea, so a large chunk of this community isn't licensed to run the local weights at all. For those users the hosted call isn't a convenience, it's the only legal door. Add that Wan 2.5 and later ship API-only, and calling a video model from inside the graph reads a lot less like a compromise than it did in 2024.

What happens when you hit Run

The node posts to https://api.acedata.cloud/minimax/videos with your ACEDATACLOUD_API_KEY as a bearer token, and - this is the part people miss - it forces the async path, so it doesn't sit there for three minutes. You get back status: pending and a task_id almost immediately. The render is now the server's problem. Drop the task_id into MiniMax H3 Retrieve Task, or wire first_media_url into MiniMax H3 Load Video once it's done.

Before the request goes anywhere, the node validates your payload against a copy of the published API contract, so a misspelled field name fails locally with Invalid parameters: does not meet the API's additionalProperties constraint instead of costing you a call. Good design, and rarer than it should be.

The fields that actually matter

content is the trap. It looks like a prompt box. It is not a prompt box - the API wants an array of content parts, and the node parses whatever you type as JSON. Type a cat surfing and you get content must contain valid JSON. What works:

[{"type":"text","text":"a cat surfing a wave at golden hour"}]

The other part types are the interesting bit, because this is what "unified context" means in practice:

[
  {"type":"text","text":"she turns and smiles at the camera"},
  {"type":"image_url","image_url":{"url":"https://example.com/her.jpg"},"role":"first_frame"},
  {"type":"video_url","video_url":{"url":"https://example.com/motion.mp4"},"role":"reference_video"},
  {"type":"audio_url","audio_url":{"url":"https://example.com/vo.mp3"},"role":"reference_audio"}
]

Text is capped at 7000 characters. Image parts take first_frame, last_frame or reference_image; video parts take reference_video (that's the V2V motion transfer people compared to Kling); audio takes reference_audio.

model has exactly one real choice, MiniMax-H3. Leave it on the (from JSON/default) sentinel and it resolves from JSON or the provider default.

resolution, duration, ratio and callback_url are plain strings - 768P or 2K, an integer 4–15, and one of adaptive/21:9/16:9/4:3/1:1/3:4/9:16. Defaults are 768P, 4 seconds, 16:9. Any non-empty widget overrides the matching key in parameters_json.

480P is listed in the contract and is a dead end: the node refuses it client-side with "The verified MiniMax H3 service supports 768P and 2K." Save yourself the confusion and don't type it.

Outputs, and how to use them

Six strings: status, task_id, media_urls_json, first_media_url, result_json, trace_id. On submit, status is usually pending and everything URL-shaped is empty - that's normal, not a failure. task_id is the one you must not lose; first_media_url is what you feed the loader later, and result_json is the response with internal routing fields and any request body scrubbed out. trace_id is your reference if you have to ask the provider what happened.

Install

cd ComfyUI/custom_nodes && git clone https://github.com/AceDataCloud/MiniMaxH3ComfyUI
python -m pip install -r requirements.txt   # requests, jsonschema - that's the lot

Or search MiniMax H3 in ComfyUI Manager and confirm publisher acedatacloud (registry ID ace-minimax-h3). Then set ACEDATACLOUD_API_KEY in the ComfyUI process environment and restart - a key pasted into the node or the workflow JSON travels with every workflow you share. The pack is new and has essentially no community footprint yet, so this is a pack where you read the source before trusting it: it's ~1000 lines of readable Python, no obfuscation, no bundled binaries.

Where people get burned

• Nothing generates after a connection drop. Paid submissions get exactly one attempt by design; read-only queries get three. If the request dies mid-flight, check your request history for an accepted task ID before queueing again, or you'll pay twice for one clip.
• Missing key throws Set ACEDATACLOUD_API_KEY in the ComfyUI process environment and restart. - that's an environment problem, not a node problem, and restarting ComfyUI is genuinely required.
• Credits. Every accepted generation bills at the current service price, scaled by duration and resolution. Query costs nothing; regenerate costs again.
• A failed render surfaces as an exception, not as a polite status: failed. Keep the task ID from the error and use Retrieve Task to see what the service says.