emaki Text to Video (Seedance 2.5)

You can't run Seedance 2.5 locally. Not on a 4090, not on a rented A100, not ever - ByteDance ships the whole Seedance line as an API product and has never opened a flagship generator's weights. So the only question this node answers is a boring one: call it from a graph, or from a browser tab?

Why the graph wins: reproducibility. The prompt lives in a workflow you can save, the clip lands in your output folder next to everything else, and you can chain an upscale or frame extraction onto the result. It's an HTTP client dressed as a node, and the costume is the point.

How it works

Prompt, model id and parameters go to emaki's generation endpoint; the node gets a run id, polls every five seconds, and when the job reports completed it downloads the MP4 into ComfyUI's output directory as emaki-<unix-timestamp>.mp4. The built-in ceiling is 900 seconds - 15 minutes - after which it gives up and tells you the job may still be finishing in your emaki library.

No local model, no VRAM math, no progress bar: it sits blocked for however many minutes the render takes. Don't assume it's frozen.

The inputs

Six required, and the first one is the only creative act:

• prompt - multiline text for the whole shot. The default ("a paper boat drifting down a rain gutter, low angle, morning light") shows the format: subject, action, camera, light.
• model - seedance_2_5 (default), seedance_2, seedance_2_fast, seedance_2_mini, minimax_h3, veo_3_1. Different vendors, one key; fast and mini are the draft tiers.
• duration - 4 to 30 seconds. Read the tooltip before you crank it: Seedance 2.5 does 4–30, Seedance 2.0 stops at 15. Asking seedance_2 for 20 seconds is a request the model can't fill.
• resolution - 480p (default), 720p, 1080p. Test at 480p. This is metered.
• aspect_ratio - 16:9 by default, plus 9:16, 1:1, 4:3, 3:4, 21:9, adaptive. Unlike the image-to-video sibling, there's no first frame to adapt to, so pick a real ratio.
• generate_audio - on by default. Seedance 2.5 produces sound natively - the gap that kept closed video models ahead of open ones for a year, and a big part of what you're paying for.

Optional: api_key, and seed, which - as everywhere in this pack - is never sent to the API. It's a cache-buster: change it to force a fresh paid run, not to reproduce one.

Outputs

Two strings: video_path (absolute path to the saved MP4) and url (the hosted copy). Both are STRING, not a video or image type, so you can't wire them into a frame-interpolation or video-merge node that expects IMAGE. Point a Load Video node at video_path instead.

This node is also an output node, which means it always executes. No cache, no free re-queues. Get your graph settled before you press play, or mute the node while you tune the prompt path.

Install

ComfyUI Manager, search ComfyUI-emaki. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/hanshs474/ComfyUI-emaki

Restart; the node appears under emaki. Nothing to install - zero declared dependencies, a standard-library client, and it reuses ComfyUI's own torch/numpy/Pillow. Grab a key at emaki.ai/settings/apikeys and paste it into api_key, or export it in the shell that starts ComfyUI:

export EMAKI_API_KEY=your_key_here

The failure modes you'll actually hit

• 401 - empty api_key and no EMAKI_API_KEY visible to the server. Exports in other terminals don't reach a running ComfyUI; restart it.
• 402 or 403 - out of credits, or the model needs a purchase on the account first. seedance_2_fast is the cheaper place to make mistakes.
• 409 - one generation per account at a time. A second tab, or a double-queued queue, trips this.
• A 15-minute deadline error - long clips at 1080p. The render usually survives the timeout and shows up in your emaki library, so check there before paying twice.
• A refusal - the content filter is inside the hosted model. Rewording is the only fix; there is nothing local to ablate.

Honest take

Text-to-video is the mode I'd use last. With no first frame the model invents composition, character and lighting at once, so you burn credits on shots that are technically fine and not what you pictured. Have a target look? Make a still first - locally or with the pack's own Generate Image node - then send it through Image to Video, where motion is the only variable left.

And keep the KB's read in mind: this is the right tool for a model you cannot run and an expensive default for one you can. Seedance 2.5 landed to the predictable "is it going to cost the earth" reception, and emaki is a Japanese-language front end to ByteDance's API. Fine for one shot; for a hundred, price out the official Seedance node.