Wan - 2.2 A14B Text to Video Turbo

Text-to-video is the mode nobody's favourite and everybody needs. Wan 2.2's T2V-A14B is open weights and genuinely good, but the honest local experience is a 25-minute round trip per clip without a speed LoRA - which is a rough way to find out your prompt is boring. This node is the turbo (accelerated) cut of the same model, hosted by KIE, and it's a reasonable way to spend a dollar instead of an evening.

Use it as the cheap iteration pass: write the prompt, run it at 480p, see whether the motion reads. Then take the one that works into your local pipeline, or just keep paying per clip if the local rig isn't there.

How it works

You give it a prompt, it gives you a video. Under the hood the node posts to KIE's marketplace endpoint /api/v1/jobs/createTask with marketplace model wan/2-2-a14b-text-to-video-turbo, then polls /api/v1/jobs/recordInfo until the job lands, downloads the MP4, and archives it under ComfyUI/output/KIE-Results/videos. What comes back on the video socket is a first-class ComfyUI VIDEO object, so anything that eats video downstream - SaveVideo, a combiner, a preview - works unchanged. No weights, no VRAM, no CUDA version argument.

Inputs that matter

prompt is the only required field, 5000 characters max. Everything else is taste:

• resolution - 720p (default) or 480p. This is the main cost dial, and 480p is the sane thing to iterate at.
• aspect_ratio - 16:9 or 9:16. Landscape or vertical; there's no square.
• enable_prompt_expansion - off by default. Flip it on for short prompts; an LLM fills in the detail while keeping your intent. It also means your prompt gets rewritten by a model you can't see, which is worth knowing when results drift.
• acceleration - none or regular, defaulting to none, with the author's own recommendation being to leave it. Faster, softer.
• seed is typed FLOAT (keep it whole for reproducibility), nsfw_checker is off by default, timeout_seconds sits at 1200 and callback_url stays empty.

Outputs

Wire video to something that saves or displays. url gives you the KIE-hosted MP4 link, all_urls_json collects every result URL, task_id is the handle you'd feed to the pack's task-status node to debug a slow job, and raw_json is the provider's unedited response. The last two sockets are credits_consumed and credits_left - after every run the node also shows a receipt on its surface, which is the nicest feature in this pack.

Install

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

Restart, and you'll find it under KIE Next / Video / Wan. The pack registers its model nodes from KIE's live documentation catalog on startup, which is why there are individual nodes per model instead of one generic "call any API" box - the README is explicit that the old generic node is gone on purpose.

Then the key, once: ComfyUI → Settings → KIE.ai Nodes Next → Connection. Paste, validate, ✅ CONNECTED, masked fingerprint displayed. Nothing about your key goes into the workflow JSON. Server installs can use KIE_API_KEY instead.

Common issues

The graph finishes and nothing appears. Add a video node. This isn't an output node - it returns a VIDEO, and if nothing consumes it you get silence.

Long queue times, then an error. timeout_seconds defaults to 20 minutes. At 720p during peak hours, raise it. A timeout doesn't refund the generation.

"Your ComfyUI build does not expose comfy_api.latest.InputImpl.VideoFromFile." Update ComfyUI - native video output needs a recent build.

Prompt expansion quietly rewrote your shot. It's off by default for a reason. Turn it on for vague prompts, off when you wrote exactly what you wanted.

Repeated near-identical clips. T2V has no first frame to anchor to, so seeds wander more than in I2V. Lock the seed between runs when you're comparing prompt variants.