MiniMax Video Generation

The MiniMax Video Generation node is the flagship of the ComfyUI-JM-MiniMax-API pack, and it's the node that answers a very specific question: what if you want Hailuo-quality video but don't have the GPU (or the patience) to run a video model locally?

Quick context: Hailuo is MiniMax's video model, and it sits in the same closed-source club as Kling, Veo, and Wan's API - the commercial video services the community compares when open-weight models don't cut it. This node is the front door to that service from inside ComfyUI. It handles text-to-video, image-to-video, and subject-referenced video, submits the job to MiniMax's cloud, and hands you a task_id to track.

How it works

The node POSTs to https://api.minimaxi.chat/v1/video_generation with your model, prompt, and settings. Any images you connect are converted from ComfyUI's tensor format into base64 JPEG and embedded in the request - the node enforces MiniMax's image rules itself (aspect ratio between 2:5 and 5:2, short side at least 300px, under 20MB). The response is a task_id, because video generation is asynchronous: the actual render happens server-side, which is why this node is always followed by Check Video Status and then Download Video.

The inputs that matter

• api_key - the pack-wide MiniMax key.
• model - the big list: T2V-01-Director and T2V-01 for text-to-video, I2V-01-Director / I2V-01 / I2V-01-live for image-to-video, S2V-01 for subject-referenced, and MiniMax-Hailuo-02 (the default, and the current best). Pick the mode, then the model.
• prompt - up to 2,000 characters. With the Director models you can embed camera-movement brackets like [左移] (pan left), [推进] (push in), [变焦拉近] (zoom in) - yes, they're in Chinese, that's the API's native vocabulary.
• prompt_optimizer - defaults to true, which means MiniMax rewrites your prompt for you. Leave it on unless you've already tuned your wording.

The optional inputs are where Hailuo's advantage shows:

• first_frame_image / last_frame_image - image inputs for I2V. last_frame_image is Hailuo-only (it generates a video ending on that frame) and doesn't work at 512P. For text-to-video, leave them empty.
• duration - 6 or 10 seconds. The 01-series is locked to 6s; Hailuo does 10s but not at 1080P.
• resolution - 512P, 768P (default), 1080P. Only applies to Hailuo; the 01-series is fixed at 720P. The node validates the combos and raises before calling the API if you pick something impossible.
• callback_url - optional; MiniMax will POST status updates there instead of you polling.

Output

One string: task_id. That's it - this node never produces video, it produces a job number. Wire it into Check Video Status.

Common issues

The mode detection is strict: pick an I2V model without connecting a first_frame_image and it refuses; pick a T2V model with an empty prompt and it refuses. Both raise clear errors.

The one that confuses everyone: error 2013 with the message "group_id can not access video 02" - that means your account isn't whitelisted for the Hailuo-02 model. The node even special-cases this message and tells you to check your MiniMax account permissions. Not a bug, just an access gate.

Install

Part of the ComfyUI-JM-MiniMax-API pack:

cd ComfyUI/custom_nodes
git clone https://github.com/synthetai/ComfyUI-JM-MiniMax-API
pip install -r requirements.txt

Or search "ComfyUI-JM-MiniMax-API" in ComfyUI Manager, then restart. Dependencies are just requests and Pillow - the video renders on MiniMax's machines, so the node itself is featherweight. Your MiniMax API key is the real requirement, and every render is a paid cloud call.