zhenzhen-hailuo-h3-max-video-lowprice（4合1）

This node runs MiniMax's Hailuo H3 Max video model on a server somewhere else and drops the finished clip back into your ComfyUI graph as if it had come off your own GPU. The trade is what you'd expect from an API wrapper: your prompt and frames leave the machine, every run is a paid cloud job, and the price tag is a lot smaller than running H3 yourself would be.

It comes from T8mars/Comfyui-zhenzhen ("zhenzhen"), a community pack that wraps cheap, closed-model APIs - Sora, Veo, Nano Banana, Seedream, and dozens more - into hundreds of ComfyUI nodes. The "2合1" in the display name means one node covers both Hailuo H3 Max flavors: hailuo-h3-max-t2v for text-to-video and hailuo-h3-max-i2v for image-to-video. You'll find it under the zhenzhen/Seedance2 Low Price tree, which is this pack's catch-all folder for its bargain-basement storefront - yes, the name says Seedance, but the whole cheap API family lives there. Don't let it throw you.

How a "generation" actually works

Nothing runs locally. The node validates your inputs, resolves your API key (more on that below), and for image-to-video first uploads your frames to the server as PNGs. Then it submits a task, polls until the cloud job finishes - you get a progress bar - and downloads the finished MP4 back to disk. One generate call, one paid cloud job, one local video file.

The image-to-video path expects image1 as the mandatory first frame and treats image2 as an optional last frame for a start-to-end transition. Pick hailuo-h3-max-t2v and the frame slots disappear entirely; the ratio picker, which only makes sense for text-to-video, shows up instead. The front-end JavaScript hides and reveals these as you flip the model dropdown, so don't panic if inputs vanish when you switch.

The inputs that matter

• model - hailuo-h3-max-t2v or hailuo-h3-max-i2v. Everything else about the UI reshapes around this.
• prompt - your actual prompt. Blank prompts are rejected before anything is charged, so no worries there.
• seconds - 5 to 15, in one-second steps. More seconds, more cost.
• resolution - this node tops out at 480P or 768P. If you were hoping for H3 Max's 2K tier, this isn't the door - it's the budget channel, and the caps are what keep it cheap.
• ratio - six fixed options (21:9 down to 9:16), shown for text-to-video only.
• image1 / image2 - IMAGE inputs, shown only in i2v mode. First frame required, last frame optional.
• api_config - wire this to the zhenzhen Low Price Settings node that holds your base_url and API key.
• skip_error - on by default it isn't; turn it on and a failed job returns an error placeholder video instead of throwing.
• seed - the tooltip says it plainly: this is an execution seed for ComfyUI's cache control only. It is not sent to the API. Leave it fixed and inputs unchanged and ComfyUI will reuse the cached result instead of charging you for a rerun.

What comes out

Four outputs. video is the downloaded MP4 as a standard ComfyUI VIDEO - preview it or save it like any other video node's output. video_url is the hosted URL if you want it elsewhere, task_id is handy for looking a job up on the zhenzhen site's task console, and response is the full JSON verdict if you're debugging.

Installing it

Install the pack once, through ComfyUI Manager (search "Comfyui-zhenzhen") or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

Then restart ComfyUI. Manager handles the dependencies, which are heavier than this node deserves - the requirements list pulls in things like opencv-python and transformers - and there are no model files to download at all. The real requirement is a paid API key from zhenzhen's storefront (the domestic China site api.seedance.nz is the default base URL and runs 7.5–8折; the overseas site ai.t8star.org is pricier), pasted into their Settings node. Their sign-ups have been intermittently closed to new users, so check the current state before you bank on it.

Where people get burned

The most common stumble is a completely empty node run that still costs credits, because someone re-executed the graph with the seed on "randomize" - every fresh run is a new paid job. Set the seed to fixed when you're iterating on something else. Timeouts don't always mean failure either; if the node gives up but the site shows the task running, pull the MP4 from the async-task console. And since the provider's own servers hiccup (the README is candid that 500s happen and a rerun usually fixes them), budget for the occasional retry. If you're outside the regions its network reaches cleanly, you'll likely need a VPN in TUN mode to even reach the API - 443 errors are the tell.