心宝视频生成-Veo

Google's Veo 3.1 has no open weights. If you want its video - and especially its native audio-with-video, the one capability the open ecosystem still hasn't matched - you have exactly two doors into a ComfyUI graph: Comfy's official Partner Nodes, or a third-party API wrapper like 心宝视频生成-Veo (XinbaoVeoVideoGenerator). This is the latter, from the comfyui-banana-li-linux pack by Li Xinbao (李心宝).

The node is a client, not a model. You give it a prompt and an API key, it packages the request, uploads any reference images to public URLs, submits to a reseller channel, polls until the clip is done, and downloads the result into the pack's XinbaoVideos folder. Every task lands in a history.toml file, and a "🎬 心宝视频任务" panel in the bottom-right corner of the UI tracks progress. Nothing runs on your GPU - your machine is just the remote control.

The inputs that matter

• prompt - required, the actual video prompt. This is a Google model, so it reads natural language well.
• banana_api_key - leave empty and it reads your key from config.ini or a global key. If you paste it into the node, don't share that workflow.
• model - veo_3_1 or veo_3_1-fast (the fast tier trades some quality for speed).
• aspect_ratio - 9:16 or 16:9 (defaults to portrait 9:16).
• batch_size - 1 to 8, with a 2-second stagger between requests. Start at 1; each clip is billed.

The image inputs are where it gets interesting. One reference image gives you image-to-video. Two images means first-frame/last-frame mode - Veo animates between your two stills. Three images switches to element-reference mode. That's the whole reason the node takes up to three image_* ports instead of one.

Outputs: video (single clip), videos (the list, for batch), and text (a STRING that carries status or error messages - in workflow-concurrency mode a failed node writes its error here instead of killing the whole graph).

Installing

Same as the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/98624017/comfyui-banana-li-linux

Restart, or use ComfyUI Manager and search "banana-li-linux". This is the Linux build - Windows users want comfyui-banana-li, macOS comfyui-banana-li-mac. The pack's requirements.txt covers requests, opencv-contrib-python, Pillow and friends.

Where people get burned

• Cost is per clip, and it adds up. Veo is one of the expensive models, and batch_size multiplies it directly. The README's own advice for the image side - keep batches small and validate before you scale up - applies double here. A session of failed prompts on a batch of 8 is an expensive lesson.
• Resolution is not in your control here, and the default disappoints. This node has no resolution input, and the Veo 3.1 API's default output is 720p. That's a real, community-hit complaint - someone ran the standard template and got 1280x720 from a model capable of FHD or 4K, wondering where their resolution went. Plan to upscale afterward, don't expect 4K out of the box.
• The route dropdown is a trap in disguise. The 线路 (route) list currently offers only 心宝❤新渠道, and the tooltip is blunt about it: the key is not universal ("Key 不通用仅为提示"). A key you bought for an older channel may simply not work here.
• Your reference images leave the machine. They're uploaded to a public URL to reach the API. That's inherent to any API wrapper, but worth remembering if your source images are sensitive.
• The security grain. This is a third-party wrapper that holds a key and phones home by design - exactly the node category the ecosystem has been burned by before. It also ships compiled .so modules rather than readable source. Read what a fresh pack does before you paste a key into it; that's not paranoia, it's the standing advice for the whole category.

If you need Veo's audio-plus-video and you're okay with metered, cloud-only generation, this gets you there with a Chinese-first UI and no Comfy credits requirement. Just budget for it, and don't expect 4K.