minimax

This node is how you run MiniMax's Hailuo video engine inside a ComfyUI graph without owning a single GPU for it. No model downloads, no 20GB checkpoints, no VRAM math. You type a prompt, the job goes up to Civitai's fleet, and a finished clip comes back down as a native VIDEO tensor. It's billed in Buzz, and it behaves exactly like Civitai's web generator - same engine, same content filter, same meter.

It's one of ~160 nodes in the official civitai-comfy-nodes pack, all generated from Civitai's Orchestration API. That matters for two reasons. First, everything here is cloud-side: the node submits a workflow, long-polls while it runs (you watch progress in the ComfyUI progress bar), and returns the result. Second, the pack is young - "early preview" in its own README - so behavior can shift between releases. Worth knowing before you build a production workflow on it.

The inputs that matter

This is a deliberately thin node, and that's the point. The only required input is prompt (multiline). Everything else is optional:

• model - a dropdown whose real choice is hailou (MiniMax's Hailuo engine; the name is just MiniMax's internal label, spelling included). Leave it empty for the default and you're fine.
• enable_prompt_enhancer - on by default. It rewrites your prompt server-side before generation. Leave it on unless you write very specific prompts and want them passed through untouched.
• source_image - an image socket. Wire a LoadImage into it for image-to-video: Hailuo animates your first frame forward instead of inventing a scene from text. The tooltip describes the underlying API field as "A URL, A DataURL or a Base64 string" - the node just converts whatever IMAGE you feed it into a data URL for you, so in the graph it behaves like any image input.
• api_config - the auth socket; see below.

Outputs and wiring

Three outputs, and you'll mostly touch one:

• video (VIDEO) - wire this into a Save Video / preview node. It drops back as a native Comfy type, so the rest of your graph (upscalers, frame extraction, a VHS-style player) treats it like any locally-generated clip.
• workflow_id (STRING) - the orchestrator job id. Handy for finding the run in Civitai's sidebar history, and for support tickets.
• raw_json (STRING) - the full workflow response, including per-currency cost. This is where you check what a run actually billed you.

Installing and authenticating

Install the pack once; every node in it shares the same setup. ComfyUI Manager → Custom Nodes Manager → search Civitai Comfy Nodes → Install, then restart. From the shell:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt   # just `requests`

Credentials resolve in this order: a wired Civitai Auth node, then the CIVITAI_API_TOKEN env var, then a stored API key or OAuth login from the Civitai sidebar. Headless/remote installs should use the env var. That's the whole dependency story - the pack's requirements file is a single line.

Common issues

• It's slow and it costs. Cloud video gen meters per job, and at launch (June 2026) the community reported long queue times - Civitai's own datacenter had just been gutted for crypto mining. A default 30-minute timeout applies; if a job times out, raise it in the Civitai Auth node (timeout_minutes) or set CIVITAI_COMFY_TIMEOUT.
• "Missing required input: prompt" - that's the node's own client-side guard catching an empty prompt before it hits the API, so the error is local and instant.
• Same filter as the web generator. Anything Civitai's moderation refuses in the browser, it refuses here. There's no local abliteration of a closed model.
• Early-preview bugs. The pack's own announcement thread carried reports of rough edges. If a node misbehaves, that's the live repo's issue tracker - the nodes are generated from the API spec, so bugs get fixed by regeneration, not by you editing node code.

Reach for this when you want Hailuo-quality clips and you'd rather spend Buzz than buy a bigger card. If you want the same idea with more dials - samplers, steps, width and height - the Wan nodes in this pack are where the tinkering lives.