Seedance 2.0 Mini

Mini is the bottom rung of the Seedance 2.0 ladder and its job is specific: find out whether your idea works before you pay flagship money for it. The structure is identical to Seedance 2.0 and 2.0 Fast - prompt, first frame, last frame, reference images, reference videos, reference audio, return_last_frame, web_search - with the resolution ceiling at 720p and 480p. Think of it as a look-development node rather than a delivery node.

The reason that matters more here than anywhere else in the pack: hosted video is the expensive part of this ecosystem, and the community's reaction when Seedance-class models arrived as API nodes was divided exactly along that line - half the room priced it and left, half pointed out that this is what a good render costs and the cheap era was always temporary. Mini is the node that lets you agree with both.

Inputs and outputs

prompt is required, multiline, documented up to 20,000 characters. The media fields are all optional native sockets: first_frame_url and last_frame_url (IMAGE), reference_image_urls (IMAGE, accepts a batch), reference_video_1/2/3 (VIDEO), reference_audio_1/2/3 (AUDIO). Same mutual exclusion as the rest of the 2.x family: KIE documents First/Last Frame and Multimodal Reference as separate scenarios, and the pack validates it locally so an invalid mixed setup fails with a useful error instead of a billed mistake. That's a genuinely good behaviour to have on the cheapest node, because the failed attempt is where beginners learn the rule.

Controls: resolution (480p, 720p), aspect_ratio (default 16:9, with adaptive available), duration (INT, 4–15 seconds or -1 to let the model choose - default 15), return_last_frame to get the final frame as an image, generate_audio (off by default, costs more on), web_search for text-to-video scenes per the docs, and nsfw_checker off by default. Plus the shared timeout_seconds and callback_url (leave that empty unless you're running a server that can receive KIE webhooks).

Outputs: video, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left. Wire video to a preview or save node; keep all_urls_json if you're chaining artifacts, and watch credits_consumed for a couple of runs so you know what your iteration loop actually costs.

Working the draft loop properly

The pattern that saves the most money: reference still in, 480p out, duration short. If you already have an approved image of your subject, start from first_frame_url - you've removed the one variable (the model's own casting) that no prompt tweak will fix, and you're testing motion and framing instead. Get that right, then move to a bigger node for the final pass, or re-run Mini at 720p if the target is social.

Don't use Mini to judge fidelity. It's the wrong tool for that, and people burn credits re-testing at 480p and concluding the model is soft. Judge motion and intent here, quality elsewhere.

Install

Manager → search KIE.ai Nodes Next (publisher id felipeosilva) → install. If it isn't listed - the README's registry publication is still described as planned rather than done - clone and install requirements:

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

Restart ComfyUI. Requirements are requests, Pillow, numpy, PyYAML; there's no model file to download and no local inference, which is the point and the catch. Windows users get a one-click .vbs installer in dist/. Then set the key once under Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key: validated against KIE before it overwrites a saved key, masked in the UI, never stored in the workflow. On servers, KIE_API_KEY in the environment takes priority.

Gotchas

Mini, Fast and the flagship are three separate nodes with three separate field sets - they're minted from three separate documentation pages, so a setting you saw on one may simply not exist here. If a node's widgets look wrong for the model, run KIE • Refresh Model Catalog.

The long pause is the render. These nodes wait on the task, holding your queue; timeout_seconds (default 1200, max 7200) is the only lever if a job outlasts your patience. Finished results are archived into output/KIE-Results/ so a restart doesn't eat them.

Everything leaves your machine. Prompt, frames, reference audio - all uploaded to KIE, all metered, and ByteDance's model applies its own content decisions no matter what nsfw_checker says. Mini being cheap doesn't make it private.