视频｜Veo 3.1 多模型视频生成

Google's Veo is the flagship of closed video models, and Veo 3.1 gets real community traction - over a hundred threads mentioning it in the Reddit corpus, mostly people wishing they could run it locally and settling for an API. The Tikpan Veo node is that API, in graph form: pick one of seven Veo 3.1 model flavors, write a video prompt, optionally feed it a first frame, last frame, or up to three components "垫图" reference images, and get a rendered video back saved to disk and exposed as a VIDEO output. No weights, no GPU, and it still costs money per second of output - the eternal API-node trade (external-api-nodes.md).

The model menu is the interesting part, because "Veo 3.1" is really a family:

• veo_3_1-lite - cheapest, fastest, the default.
• veo_3_1-lite-4K / veo_3_1-fast-4K - 4K clarity at lite/fast speed, pricier.
• veo3.1-pro - top quality, slowest.
• veo3.1-fast-components / veo_3_1-components-4K / veo_3_1-fast-components-4K - the components models, which are the ones that accept the reference "垫图" inputs.

How it works

Submit a text-prompt job to Tikpan's Veo endpoint (/v1/videos or the legacy /v1/video/create depending on model), then poll until the render completes - the node waits in a loop, printing a "云端渲染中" progress line every 15 seconds. On completion it downloads the video locally and returns the path plus a VIDEO tensor. The optional 参考图_首帧 and 参考图_尾帧 set the first/last frames; the three 垫图_1..3 inputs are visual references - and the tooltip is explicit that only the components model series honors them. Feed 垫图 to a non-components model and they're simply ignored. The 比例 dropdown is just 16:9 and 9:16 - a deliberate trim for TikTok-format work.

Inputs and outputs that matter

API_密钥, Veo专属提示词 (English recommended, the more specific the better), 模型选择, and 比例. The optional 校验HTTPS证书 (default on) is a local networking toggle for weird proxies - the tooltip notes it doesn't get sent upstream. Outputs: 📁本地保存路径, 🏷️任务ID, 🔗视频云端直链 (handy if you want to hand a URL to a video-pipeline step), 📄完整日志, and 🎬视频输出 (VIDEO).

Installing

Pack install as usual - Manager search "Tikpan", or git clone https://github.com/htrert/ComfyUI-Tikpan-Pro into custom_nodes, restart. Key from tikpan.com. Nothing local to download.

Common issues

• Video takes minutes, not seconds. Veo renders on Google's side; 4K and pro flavors especially. The node polls with no hard max wait visible in the required fields - if you're using this in a batch, consider submitting and collecting via the Task Fetcher node instead of blocking the queue.
• 垫图 inputs doing nothing. You picked a non-components model. Switch to one of the components flavors - the tooltips call this out precisely.
• Portrait vs landscape mismatch. Only 16:9 and 9:16 are offered, and the model family itself is where the quality choice lives. For TikTok vertical, 9:16 + fast/lite is the sensible default; don't burn pro-tier tokens on short loops.
• Seed is hidden. The node hides the seed field because it isn't a confirmed reproducible parameter through this relay - don't expect deterministic re-runs. This is the pack being honest rather than exposing a knob that doesn't work.
• Cost. "A Veo 3.1 second of 4K video" is priced individually per the ecosystem docs - a 4K clip adds up fast (external-api-nodes.md). Start with lite.