Hailuo H3 Max 视频生成（4 合 1）

The name is doing two jobs here, and it helps to know both. "Hailuo" is MiniMax's video brand - the same lab behind the H3 open-weights release - and "Max" is the newer, hosted step up the line. Hailuo H3 Max 视频生成（2 合 1） is a ComfyUI node that sends your prompt to that model over the internet and drops the finished MP4 back onto your canvas. No checkpoints to download, no VRAM budget, no GPU at all. That's the whole pitch, and for a model that only exists as a paid API, it's the only door.

This node is one of dozens in the ComfyUI_Seedance pack (publisher t8star) - one author's giant front-end for a paid gateway at api.seedance.nz that resells closed Chinese video and image models. It was added in pack v0.11.0 as a standalone 2-in-1 rather than bolting two more options onto the existing Hailuo H3 node, so old H3 workflows don't move.

How it works

Behind the scenes this is an API wrapper, not a sampler. Pick a model, write a prompt, hit run, and the node uploads any frames, POSTs to the /v1/videos endpoint, polls the remote task until it succeeds, then downloads and verifies the result before handing you a ComfyUI VIDEO. The whole pipeline - upload, submit, poll, download - is baked in, with a progress bar on top.

The "2 合 1" is where the model-specific weirdness lives. There are exactly two models in the dropdown:

• hailuo-h3-max-t2v - text-to-video. It requires a fixed aspect ratio (six choices: 21:9 down to 9:16, default 16:9). No "adaptive" escape hatch here.
• hailuo-h3-max-i2v - image-to-video with a first frame and an optional last frame. It does not send a ratio at all; the output follows whatever your input frame looks like. Connect image1 or it'll refuse before uploading anything.

Both models share the same limits: prompt is required (1–20,480 chars), duration 5–15 seconds, resolution only 480P or 768P. Note that cap - the older H3 node in this pack goes up to 2K, but H3 Max through this gateway tops out at 768P.

The inputs and outputs that matter

You mostly touch three things: model, prompt, and - for image-to-video - image1 (plus optional image2 as the ending frame). seconds and resolution default to a cheap 5s at 480P, which is exactly where you want to start before spending real money.

Three extras are worth knowing:

• api_config - wire in the pack's Seedance API Config node here. Skip it and the node falls back to the SEEDANCE_API_KEY environment variable.
• seed - and here's a genuinely nice detail for an API node: H3 Max has no documented seed support, so this value is not sent to the model. It's pure ComfyUI cache control. Leave it fixed and unchanged inputs reuse the cached result instead of burning another paid call; randomize forces a fresh generation.
• skip_error - on failure, output a placeholder error video so a batch doesn't die on one bad lane.

Outputs are the standard four from this pack: video (the real VIDEO, wire it into Save Video / preview), video_url (the direct link - results can expire, so save what you like), task_id, and response (full JSON if you want to poke at it).

Installing it

The pack installs like any ComfyUI custom node. Easiest is Manager: search ComfyUI Seedance or seedance, install, restart. From the CLI:

comfy node install seedance

Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/ComfyUI_Seedance
cd ../..
python -m pip install -r custom_nodes/ComfyUI_Seedance/requirements.txt

That requirements file is a single line - requests>=2.28. No torch extras, no model weights, nothing heavy. Restart and you'll find the node under the Seedance category.

First run and the honest caveats

You need a token from the gateway console (api.seedance.nz/console), set it in the config node, as SEEDANCE_API_KEY, or in a config/.env inside the plugin folder. One thing worth flagging: the README leads with affiliate links to that signup page and to RunningHub. You don't need to click them - any valid key for your configured base_url works, and keep it out of workflows you share. This whole category is "a node that phones home with your key," so the usual rule applies: only install widely-used packs, and don't paste keys into public graphs.

If a download stalls mid-way, the node quietly retries - it has an 8-second connect / 180-second total cap and falls back to a system curl channel, then a fresh session, and it validates the MP4 header before declaring victory. Missing-key errors and a strict "connect image1 for I2V" validation are the two most common walls, and both fail before anything hits the API, so they cost you nothing but a console message.

The one thing you won't find yet is community takes on H3 Max - as of writing it's brand new, with zero Reddit footprint and zero search impressions. You're early. That's fine: start at 480P, confirm your key round-trips, then decide if this newer model line earns a spot next to your local video nodes.