Seedance 2.0 Fast

If you're going to spend money on hosted video, this is the node I'd start with. Seedance 2.0 Fast keeps the interesting half of the flagship's feature set - first frame, last frame, reference images, up to three reference videos, up to three reference audio clips, return_last_frame, web_search - and gives up the top of the resolution range: 480p and 720p only, no 1080p, no 4K. For iterating on a shot, that's not much of a loss, and the model still understands what you meant, which is the only thing a draft pass needs to prove.

That's the practical read on the whole "Seedance is expensive" argument. The people running these nodes successfully aren't queuing one perfect 4K clip; they're burning cheap drafts at 480p until the prompt, framing and motion are right, then spending once at the quality they're shipping. Fast is the node for the first 90% of that process.

Inputs and outputs

prompt is the only required field - multiline, docs say up to 20,000 characters. Then the media sockets, all native: first_frame_url and last_frame_url (IMAGE), reference_image_urls (IMAGE, batches fine), reference_video_1/2/3 (VIDEO), reference_audio_1/2/3 (AUDIO). Same rule as the rest of the Seedance 2 family: First/Last Frame mode and Multimodal Reference mode are mutually exclusive per KIE's docs, and the pack validates that locally so a mixed workflow errors out before it costs you a call.

resolution (480p/720p), aspect_ratio (default 16:9, with adaptive as an option), and duration (INT, default 15, documented as 4–15 seconds or -1 to let the model choose). return_last_frame gives you the final frame as an image for chaining. generate_audio off by default and costlier when on. web_search exists but the docs restrict it to text-to-video scenes. nsfw_checker defaults off, which switches off KIE's check - the model's own moderation is a separate matter you can't reach.

Outputs are the pack's video signature: video, url, all_urls_json, task_id, raw_json, and credits_consumed / credits_left. Those last two, plus the credits-spent and balance display that the node shows after each run, are how you keep a metered workflow honest.

One quirk worth naming

The input list on this node contains two sets of reference-video sockets: reference_video_1/2/3 and a second trio named reference_video_urls _1/2/3. That's the pack's node factory surfacing both names KIE's documentation uses for the same field rather than a feature - the parameter mapping accepts either, so pick one set and ignore the other. It looks like a bug report and it isn't one. It is, however, a good illustration of what generated nodes cost you: they track the docs that closely, warts included.

Install

Manager → search KIE.ai Nodes Next (publisher id felipeosilva) → install. If Manager finds nothing, the README describes registry publishing as prepared rather than done:

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. Dependencies are requests, Pillow, numpy and PyYAML - nothing heavy, no model downloads, because all inference happens on KIE's servers. There's a Windows one-click .vbs installer in dist/ and a full uninstaller beside it. Set the key once in Settings → KIE.ai Nodes Next → Connection; it's validated before it replaces a saved one and only a masked fingerprint is shown. KIE_API_KEY takes priority for headless installs.

Gotchas

Don't assume Fast means the same knobs as the other 2.x nodes. The three are generated from three separate documentation pages, so their fields differ - Fast and Mini cap at 720p, the flagship adds 1080p and 4K, and the V1-era Pro Fast node drops things like camera lock and seed entirely.

Generation is synchronous, so a long clip occupies your queue. timeout_seconds defaults to 1200 and goes to 7200; a queue that looks frozen is usually a video still rendering.

Nothing here is local. Your prompt, frames and reference media go to KIE, every clip is metered, and ByteDance's content policy applies regardless of the nsfw_checker setting - there are no weights in your install to patch. That's the price of a model you can't download, and it's worth paying deliberately rather than by accident.