Hailuo Pro Text to Video

What it is

The pure prompt-to-video endpoint of MiniMax's Hailuo 02 Pro tier, at KIE Next / Video / Hailuo. No image sockets, no duration picker, no resolution menu - you write a shot and it renders one. It's the simplest node in the pack's video folder, and the one that most rewards writing well.

Where it's genuinely useful: storyboard beats, transitions, B-roll, "I need six seconds of something moving for this edit". Hailuo's closed service is one of the models people hold up when comparing against local video generation - the honest split is that local Wan or LTX costs nothing per run but eats an afternoon of tuning, and this costs a credit and takes a minute. Use the host to explore, use your card to refine.

How it works

The prompt goes to KIE's create-task endpoint; the node polls the task until it completes (up to timeout_seconds, default 1200 - remember these are video jobs, they take real time), downloads the MP4, and returns a native ComfyUI VIDEO. Nothing local, nothing uploaded, nothing to configure beyond the prompt and one flag.

The inputs

Two, really.

prompt (required) - up to 1500 characters. The author's placeholder is a wall of camera direction ("high top angle wide mid close-up tracking shot, flying very fast two meters high…"), which is a fair hint at what this model wants. Motion, camera, subject, environment - in that register. Hailuo does not need art-historical adjectives; it needs geography and movement. If your prompt reads like a still image description, you'll get a still image that twitches.

prompt_optimizer (default true) - the one switch worth understanding. On, the service rewrites your input before generation, which rescues short or sloppy prompts and can improve results on vague ones. Off, your text is used as written, which is what you want once you've learned what the model likes and you're doing controlled comparisons. If you're A/B-ing prompt phrasings, turn it off, or you're testing the optimizer instead of your prompt.

Also present: nsfw_checker (default false, with the vendor's own caveat that disabling filtering means results come back as produced and nothing is guaranteed to be screened), timeout_seconds (1200, range 30–7200), and callback_url - leave that empty for normal ComfyUI use, it's for servers receiving KIE webhooks.

Outputs

video is the clip, as a ComfyUI VIDEO - feed SaveVideo, a preview node, or a muxer. url is the primary hosted result and all_urls_json carries every URL the task returned. task_id is the job handle: keep it if you plan to look the job up later with the pack's task-detail utility, or to chain into anything that takes a KIE task reference. raw_json is the provider's full response and the place to look when a task fails, since the node surfaces KIE's own failure message rather than a generic error. credits_consumed and credits_left report the cost and your balance, and are also printed on the node itself after a successful run - the simplest way to watch video spending without opening a billing page.

Install

Manager → KIE.ai Nodes Next; manually, clone the repo into ComfyUI/custom_nodes and install its requirements.txt in ComfyUI's Python environment:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
python -m pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

requests, Pillow, numpy, PyYAML. No weights, no VRAM. Restart, then save your KIE.ai key once under Settings → KIE.ai Nodes Next → Connection (validated before it replaces anything, masked afterwards, and KIE_API_KEY for headless). The node menu is generated from KIE's docs index on first launch and refreshed about daily, with one frontend reload during the first sync.

Gotchas

• This is the priciest thing in the folder per run. Text-to-video at Pro quality with no draft mode - nothing here is cheap, so write the prompt before you queue it, not after.
• No duration or resolution control. Fixed by the tier. If you need those knobs, use the standard 02 text-to-video node instead; the pack ships it as a separate node.
• prompt_optimizer on by default means your prompt isn't your prompt. Useful, and worth turning off when you're testing.
• 1500 characters. The provider truncates; don't paste a screenplay.
• VIDEO outputs need a current ComfyUI build - the pack uses comfy_api's file-backed video helper and tells you explicitly to update ComfyUI if it's missing.
• No key → No KIE API key configured. That's the one-time Settings step, not your graph.