zhenzhen-hailuo-h3-video-lowprice

MiniMax H3 is the omni-modal one - the video model that treats text, images, video and audio as a single input context and can spit out clips with native sound, instead of bolting an audio pass on after the fact. This node exposes almost all of that surface area at once, which makes it the most flexible node in the Zhenzhen lowprice family and also the easiest to get lost in.

The hailuo-h3-* model list behind it is why you'd reach for this over the plain Hailuo 2.3 node: t2v, i2v, multi (mixed image/video/audio references), and global- variants of each - the "global" line being the looser-review, character/multi-reference tier. If your clip needs a reference video and a voice clip and a couple of stills, there isn't another node in this pack that accepts all of them at once.

How it works

Standard lowprice mechanics: an api_config input (ZHENZHEN_SEEDANCE2_CONFIG type) comes from the pack's Comfly_seedance2_low_price_settings node, which holds the base URL (api.seedance.nz, the cheaper domestic site) and your API key. The node submits, polls, downloads, and hands back a VIDEO. What's unusual is the input spread: up to 9 images (image1–image9), 3 videos (video1–video3) and 3 audio clips (audio1–audio3). Those map to H3's omni-modal context - for multi and global-multi models, a reference video drives motion while a reference audio can drive the sound.

Inputs that matter

• model - the required enum. Start with hailuo-h3-t2v for a text-only test; move to i2v with an image1, or multi when you need to mix media types.
• prompt - describe the scene. H3 is genuinely multimodal about it, so the prompt can reference what you've wired in.
• seconds - 5 through 15, string enum. Longer = pricier; 5s is the cheap sanity check.
• resolution - 768P or 2K. ratio - adaptive plus the usual 16:9/4:3/1:1/3:4/9:16/21:9.
• The image* / video* / audio* inputs - only connect what the model variant uses. A t2v run doesn't need your 9 images wired up.

Outputs: video, video_url, task_id, response - standard pack shape, and task_id is your backdoor if ComfyUI times out but the job finished (check the site's async-task page).

Installing it

ComfyUI Manager → search Comfyui-zhenzhen → install, or:

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

Restart, drop in the settings node with your key, wire its api_config across. No weights to download; the generation happens on MiniMax's side and you're billed per run.

Common issues

• Wrong model for the inputs - the biggest trap. t2v models will silently ignore your reference videos, and multi models expect at least one reference beyond the prompt. Match model to what you wired.
• References not "sticking" - H3 is better than most at keeping a subject from reference media, but "better" isn't "perfect"; over-stuffing the inputs is a real cost and consistency risk, so trim to the strongest 2–3 references.
• 2K being slow/expensive - expected, not a bug. Drop to 768P for iteration, keep 2K for finals.
• The standard pack gotchas - empty key in the settings node, overseas-server 443 errors (VPN in TUN mode per the README), and 500s that are the vendor's problem, not yours (re-run).

One aside: H3's open weights exist and you can run the 33B locally - but it wants serious VRAM and an audio pipeline. This node is the "I don't have the hardware" path, and it's honest about being billed per clip.