Veo 3.1 (Replicate)

Veo 3.1 is Google's closed video model and the closest thing this ecosystem has to a quality benchmark you can't download. The community says the open models caught the still-image gap, but video is a different story - and Veo's one genuinely unmatched capability is native audio: it generates synchronized sound, dialogue, and music with the video, not bolted on afterward. As of mid-2026 no open model quite does that. This node drops Veo into your ComfyUI graph for the price of a per-call API fee.

Set the expectations the same way you would for any closed video API: the prompt and any reference images leave your machine, Google's content policy decides what's allowed, and the output is a file path - not a tensor you can keep editing in-graph.

How it works

Standard Replicate Select engine around google/veo-3.1. The node waits for the worker, downloads the MP4, saves it to output/video/, and returns the video_path STRING. It's an output node, so the clip shows up in the UI history. generate_audio defaults to true - which is the whole reason you'd pick Veo over a silent video model.

The inputs that matter

• prompt (required) - describe the video. Veo reads detailed scene descriptions well.
• duration - an enum, not a slider: 4, 6, or 8 seconds. No 5s option; that's a Veo quirk you'll just live with.
• aspect_ratio - 16:9 or 9:16 only.
• resolution - 720p or 1080p (default). 1080p is the better deal for the money if your prompt deserves it.
• image - a first frame to start from. Ideal inputs are 1280x720 (or 720x1280) matching your aspect ratio.
• last_frame - an ending image; combined with an input image this creates a transition between the two.
• reference_images - 1–3 images for subject-consistent generation (reference-to-video, R2V). Two constraints worth remembering: reference images only work with 16:9 and 8 seconds, and when you use them, last_frame is ignored.
• negative_prompt - what to exclude.
• generate_audio - on by default. This is the headline feature; leave it on.
• seed - set for reproducibility, omit for random.

Output: video_path, a STRING.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", restart. Dependency: replicate>=1.0.7. Auth as an environment variable:

export REPLICATE_API_TOKEN=r8_...

Where people get burned

• R2V constraints are hard requirements. Reference images + 9:16, or reference images + 6 seconds, and the config is just wrong. Stick to 16:9 @ 8s when you want subject consistency.
• last_frame gets silently dropped with references. Not an error - the model just ignores it. If you set both, you've paid for a feature you're not getting.
• Video is where sessions get expensive fast. A few takes at 1080p, with audio, add up quicker than image generation ever will. Draft with 720p and shorter durations.
• The path-not-tensor thing. video_path is a STRING for downstream use; the actual clip is the file in your output folder. If you want frames in the graph, load the file with a video node yourself.