BYOKey Google Veo Video

Veo is the model the video community measures everything else against. No open weights exist, so the only way to use it is to call it - and this node is the bring-your-own-key call, straight to Google's generativelanguage.googleapis.com with your Gemini API key. Veo 3 added native audio generation, which the KB flags as the one capability open-source video still hasn't matched, so if you need sound baked into the clip, this is the node you want.

How it works. The submit-then-poll pattern this pack is built on: POST your job to Google with your key, poll every poll_interval seconds until the operation completes, download the mp4, and wrap it in ComfyUI's native VIDEO type. An optional image input lets you steer with a reference frame (Veo 3's i2v is genuinely good). Inputs are images sent as inline base64 data URIs.

The inputs that matter - and the fiddly matrix:

• model - six options spanning two generations: veo-2.0-generate-001, plus the 3.x family (veo-3.1-generate, veo-3.1-fast-generate, veo-3.1-lite, veo-3.0-generate-001, veo-3.0-fast-generate-001). This is where most people get lost, because the valid combinations below depend on it.
• resolution - 720p, 1080p, or 4k - but 4K only works on the 3.1 models, and not on veo-3.1-lite or anything 3.0. The node raises a hard error if you ask for 4K on a model that can't do it.
• duration_seconds - 4–8, stepped. Veo 2 allows 5–8; Veo 3 is 4–8 in steps of 2. Check against the tooltip when you switch models.
• enhance_prompt - on by default and forced on for Veo 3 models. Google rewrites your prompt to "make it better." If you want your exact wording sent, you need a Veo 2 model; with Veo 3 you don't get a choice.
• negative_prompt - genuinely supported here, which is rarer in video APIs than you'd think. Use it.
• person_generation - ALLOW or BLOCK, Google's explicit gate for generating people. If you get refusals on a people-heavy prompt, this is the first switch to check.
• generate_audio - off by default; Veo 3 only. Flip it on and the clip comes with synchronized sound, dialogue, the works.
• aspect_ratio - 16:9 or 9:16.
• seed - 0 for random; Veo honors it otherwise.
• image - optional reference frame for i2v steering.

poll_interval and max_wait are the babysitters; the 10-minute cap is plenty.

Output: one output (VIDEO) socket.

Install. Shared pack, one clone:

cd ComfyUI/custom_nodes
git clone https://github.com/MeteorAndy/comfyui-byokey.git
# restart ComfyUI; it's under api/byokey/video

No model downloads, no extra dependencies beyond ComfyUI's bundled torch, aiohttp, PIL, numpy and av.

Gotchas. The model/resolution/duration/enhance matrix is the trap - read the tooltips when you switch models, and know that 4K demands a 3.1 non-lite model. Jobs are the slowest in the pack; a 4K 3.1 clip can take a while, so keep max_wait generous. And the pack is young with no community footprint - skim the source before feeding it a Google key.