Generate Veo3.1 Video · Veo 3 fast

If you're going to burn credits learning how Veo behaves, do it here. veo3_fast is the middle tier of the three Generate Veo3.1 nodes, and it's the one that supports every generation mode including reference-to-video - which makes it both the most flexible and the best value in the pack. Quality tier is for the final render. Lite is for previz. This is the one you actually work in.

The mode support is the practical difference people miss. Veo's reference mode - one to three images that the model treats as subject references rather than literal first frames - is only available on fast and lite, not on the quality model. If your workflow is "hold this character across shots", fast is where that lives.

The mechanism

The node posts the documented Veo3.1 generation request with model: "veo3_fast" fixed, gets a task ID back, polls KIE's task record (2.5s to 15s backoff) until it reports success, downloads the video, and returns it as a native ComfyUI VIDEO. Any images you connect are uploaded to KIE first and their URLs placed in the request. It also returns the task_id as an output, which is what the Extend Veo3.1 nodes want if the shot needs more than eight seconds.

One implementation detail with real consequences: for 1080p and 4K renders the final URL lands in a different field than the 720p one, and the pack sweeps the whole response for every URL, preferring the ones that look like video files. In practice you get the high-res file. If something ever looks soft, check all_urls_json before blaming the model.

Inputs

• prompt - required. Write for motion: who does what, how the camera moves, what the light does. Detailed beats vague, but don't write a novel.
• image_urls - optional. In FIRST_AND_LAST_FRAMES_2_VIDEO mode: one image animates from it, two generate the transition between them. In REFERENCE_2_VIDEO mode: one to three reference images.
• generation_type - REFERENCE_2_VIDEO is the default and is supported here, unlike on the quality node. Switch to TEXT_2_VIDEO when you have no images, or the frame modes when you do.
• resolution - 720p default, 1080p, 4k. Change this at render time, not at draft time.
• duration - 4, 6, 8. Remember reference mode only supports 8 seconds; asking for 4 with references is a rejected request, not a short clip.
• aspect_ratio - 16:9, 9:16, Auto.
• watermark - ships as MyBrand. Clear it unless you want that text on the video.
• enable_translation - on by default; converts the prompt to English first. Turn it off if your prompt is already English and you care about exact wording.
• enable_fallback - deprecated, leave it alone.
• timeout_seconds, callback_url - 1200s; empty callback.

Outputs: video, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left.

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

Manager: search "KIE.ai Nodes Next". Dependencies are requests, Pillow, numpy, PyYAML, and there is no local model of any kind - which is why a 4GB laptop can run this node. Paste your KIE key once at Settings → KIE.ai Nodes Next → Connection; the pack validates it against KIE before saving and never writes the full key into the workflow.

Practical warnings

Wire the video output into KIE • Preview Video, KIE • Save Video, or any node that takes VIDEO. This node isn't an output node, so an unwired result means it simply doesn't execute - no error, no charge, no clip. Beginners lose an hour to that.

An old ComfyUI build will refuse to make a VIDEO object at all and tell you to update, because the pack relies on ComfyUI's newer video interface. Update rather than trying to patch around it.

And the cost model, again, because it's the part that bites: drafting at fast is cheap relative to quality, not cheap. Three failed takes at 1080p is a real amount of money. Draft at 720p, lock the prompt, then re-render the keeper.