- V1 Lite Text to Video

Text to video with nothing but a prompt is still the demo everyone wants and the workflow nobody actually ships - which is fine, because V1 Lite is cheap enough to be a sketching tool. Type an idea, get five seconds, find out whether the idea reads. No image to prepare, no reference sockets to reason about, no mutual-exclusion rules to memorize. Just a prompt and a small set of dials.

The one thing I'd nag about up front: this node's aspect_ratio includes 9:21, which is a display ratio you'll hit by accident on a dropdown scroll and won't notice until you've paid for a very tall video. Set it deliberately before you queue.

Inputs and outputs

prompt is required - multiline, and the shipped default is a good example of the house style: "Wide-angle shot: a serene sailing boat gently sways in the harbor at dawn... The camera slowly pans across the scene." Name the shot and the camera move inside the prompt; this node has a camera_fixed switch but no native camera parameters, so text is how you direct it. Where a movement matters, one clear instruction beats three competing ones.

Optional inputs: aspect_ratio (default 16:9, also 4:3, 1:1, 3:4, 9:16, 9:21), resolution (480p, 720p, 1080p - 480p for drafts, 720p for the balance), duration (5 or 10 seconds, discrete), camera_fixed to lock the camera off, seed (INT, default a fixed value, -1 for random), enable_safety_checker (default true - the docs note it's always on in KIE's playground and only switchable via the API), nsfw_checker (default false), plus timeout_seconds and callback_url.

Outputs: video, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left. The last two, and the credits-spent/left receipt the node shows after each run, are what make a metered tool tolerable - you see the price of an idea at the moment you spend it.

Draft-then-commit, the only workflow that makes sense

Everything about this node argues for cheap iterations: resolution at 480p, duration at 5, and a prompt short enough to rewrite five times in a session. The advice that circulates among people using hosted video nodes is exactly this - refine at the lowest resolution, commit at the one you're delivering - and it's doubly true here, because you have no reference frame anchoring the result. Half of text-to-video is casting: if the model's choice of subject doesn't match the picture in your head, a prompt tweak costs far less than a 1080p ten-second re-roll.

When you finally get one you like, keep the seed and set the widget's control-after-generate behaviour so the value is used before generation rather than re-rolled after it - otherwise "reproduce this" gives you a different clip and you'll blame the model.

Install

Manager → search KIE.ai Nodes Next (publisher id felipeosilva) → install. If it isn't listed, the README still frames registry publication 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. Requirements are requests, Pillow, numpy and PyYAML - no ffmpeg, no local weights, nothing to download, because the generation happens on KIE's side. Results are archived under output/KIE-Results/ so a restart doesn't wipe a finished clip. Windows users can run the .vbs one-click installer in dist/. Set the API key once under Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key: validated before it replaces a saved key, masked in the UI, never embedded in workflow JSON. KIE_API_KEY wins for headless installs.

Gotchas

Turn the queue off in your head. The node waits synchronously for the task; a 10-second 1080p clip can hold your ComfyUI queue for a while, and timeout_seconds (default 1200, max 7200) is your only lever.

ByteDance's video line is API-only. V1 is the older, cheaper tier of a closed model family whose flagships have never been released as weights, so there's no local fallback to switch to when the API is slow - your local options are the open video models you already know about, and they're a different animal.

Prompt language is not free of consequence. A text-to-video prompt with no shot language gives you a moving wallpaper. Name the shot size and one camera move, or hand the writing to KIE • Camera Director and reuse the same phrasing across shots so your clips look related.

Your inputs leave the machine. Every call is metered and your prompt goes to KIE. Same deal as every hosted model: the capability you can't have locally, at per-call prices.