Wan 2.6 - Text to Video

Wan 2.6 landed in December 2025 as a commercial release, and if you were hoping for open weights, the answer two years of following this series has taught everyone is no. There's no download, no checkpoint to swap in. The only way to use it is through an API - KIE's, in this case, with a ComfyUI wrapper that takes a prompt and gives you a clip.

Text-to-video is the mode with the least control, so use it for what it's good at: setting a scene, testing a motion idea, generating b-roll you'd otherwise shoot. For anything with a specific actor or a specific look, go I2V and control frame one.

How it works

prompt is the only required input (Chinese and English, minimum 1 character, max 5000). Optional: duration (5, 10, 15), resolution (1080p default, 720p), multi_shots, nsfw_checker, timeout_seconds, callback_url.

The node submits to KIE's marketplace at /api/v1/jobs/createTask using model wan/2-6-text-to-video, then polls /api/v1/jobs/recordInfo until the job reports success or fail - the graph sits and waits, it doesn't return early. On success it downloads the MP4 to ComfyUI/output/KIE-Results/videos and your video socket emits a proper ComfyUI VIDEO. Your machine contributes nothing but an HTTP request.

Inputs that matter

• prompt - describe the shot and the motion. The shipped default is a hyperrealistic ASMR bit with a knitted knife; the model handles a small narrative, not just a noun.
• duration - 5, 10 or 15. Fifteen seconds at 1080p is the expensive end of this whole folder.
• resolution - the primary cost lever, 1080p by default.
• multi_shots - off. On, you get several shots with transitions rather than one take. Genuinely useful for a scene; fatal if you wanted a continuous camera move.

No seed field on this node, so a specific render isn't reproducible - save the outputs you like and let the url / all_urls_json sockets keep the links.

Outputs

video for anything downstream. url and all_urls_json for the hosted artefacts, task_id for tracking, raw_json for the untouched provider response, and credits_consumed / credits_left plus the on-node credits receipt for the accounting. That last pair is the reason to run this in a graph rather than a web playground: you see what each take cost while you're still in the flow.

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 ComfyUI, then find it in KIE Next / Video / Wan. Nothing downloads beyond the pack: no weights, no VAE, no model folder. Dependencies are requests, Pillow, numpy, PyYAML.

Key: Settings → KIE.ai Nodes Next → Connection, paste, validated, ✅ CONNECTED, masked fingerprint. KIE_API_KEY in the environment is the headless path and takes priority over the saved key. KIE is pre-paid credits with no subscription - exactly the shape you want when the whole point is testing a model you can't run.

Common issues

You keep paying for takes you delete. KIE charges per generation. Do duration and resolution experiments at 720p/5s, then re-run the winner at 1080p.

Job times out. Default timeout_seconds is 1200. Long clips at 1080p can queue past that; raise the value rather than killing the run.

Repeated prompts drift between runs. No seed socket here, so motion wanders. If you need the same clip twice, keep the file.

Nothing appears after a green run. Add a save or preview node. It returns a VIDEO; it isn't an output node.

comfy_api.latest error. Update ComfyUI - native video output needs a current build.