心宝♥视频生成（统一模型）

XinbaoUnifiedVideoGenerator - 心宝♥视频生成（统一模型） - is a remote control, not a generator. There's no checkpoint, no VAE, no 14B anything. You pick a model from a dropdown, type a prompt, and your request goes out to a commercial video API. Locally, a 5-second clip is the practical ceiling and 15–40 minutes on a 4090 is normal; here you ask for 10 seconds at 1080p and wait on someone else's rack. That's the whole trade: seconds instead of hours, and a bill instead of your GPU.

It replaced the author's separate Sora/Veo/Doubao facade nodes - three nodes for three families - and pulls the model list from a registry inside the pack (XinbaoVideos/core/model_mapping.py).

How it works

The node assembles a request, submits it to the pack's own relay (the banana_route widget has exactly one option in this build - 心宝❤新渠道 - so the route/widget pairing is cosmetic for now), and gets back a task ID. Anything you wired in - reference images, a reference video, a reference audio clip - gets uploaded first, which is why reference media shows up under "按连接顺序提交": order matters, and the node sends them in that order.

Then it polls. A background task manager (XinbaoVideos/core/video_task_manager.py) ticks every couple of seconds, re-checks the provider about every 15, backs off on errors, then downloads the finished MP4 with parallel range requests into ComfyUI/output/video_tasks/. The 心宝任务中心 panel bottom-right of the ComfyUI UI can re-poll, retry a download and open that folder - and it's where your key goes: stored in a local, gitignored config.ini, shown back as asterisks, never baked into the workflow JSON.

One honest note: every model in that dropdown is closed-weights (Seedance is ByteDance's API-only line, MiniMax H3 the Hailuo service's frontier model), so nothing here is local and every prompt and reference image leaves your machine. You also can't read the shipped code - it's compiled Linux .so, not Python - and a node that holds a credential and calls the network by design is the shape of the one node in this ecosystem already weaponized. Judge the pack on reputation.

Inputs you actually set

prompt is required; so is model_type, whose ten options are labelled in the UI as friendly names - 全能视频S2 / S2 Fast / S2.5 (the Seedance 2.x family), SD2.5pro特价 and the ty- prefixed entries (cheaper channel lines, priced down accordingly), MiniMax H3, Wan3.0, and omni.

duration runs 4–30 seconds and resolution offers 480p/720p/768p/1080p/2K, but not free-form: the pack validates each request against a per-model spec and refuses a resolution, aspect ratio or duration the chosen model doesn't offer. aspect_ratio defaults to 16:9 (also 9:16, 4:3, 1:1, 3:4, 21:9). generate_audio defaults on where the model supports sound; omni_seconds (4/6/8/10) only matters for omni. Leave banana_api_key blank and set the key in the task centre.

Three switches matter. 启用工作流并发 (default off) lets several video nodes fire at once, and when it's on a failure blocks just that node instead of killing the run. 仅提交不等待 returns immediately with no savable video this run - the task centre tracks it. bypass_proxy / disable_ssl are for broken networks only; turning off certificate checks is how your key walks out the door.

Outputs are video (wire it into a core Save Video / preview node), videos (the list form) and text, which is your status and error channel - that's where a failure message lands when concurrency is on.

Installing it

ComfyUI Manager → search Banana Gemini Image Generator, or the by-hand version:

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

Use the -linux URL. The mirror README still tells you to clone the Windows repo, and the pack's bootstrap checks pyproject.toml and prints a red "Wrong repository for your platform" banner if you mix them up. Restart ComfyUI after cloning.

Dependencies are lighter than the README claims: requirements.txt is three lines (requests, Pillow, toml on Python <3.11) - the torch/opencv/transformers list is the legacy image-node stack, and the video path's aiohttp is already a ComfyUI dep. The binaries are prebuilt ELF x86-64, so this build won't import on an ARM box.

Where people get burned

• "Missing binary files", or no node at all. Wrong-platform repo, or the .so files aren't there. Check the ComfyUI console for the bootstrap banner.
• "缺少 Banana API Key". The node reads the key from the task centre; a blank node field is correct, not a bug.
• Text output, no video. Either 仅提交不等待 is on, or the download hasn't finished - open 心宝任务中心 and check the task, it usually has the file already.
• Model/parameter rejections. Messages like "不支持分辨率…" mean that model doesn't do what you asked. Switch the model rather than fighting the widget.
• Cost creep. Every call is metered, including retries, and a 10-second 1080p clip is not a cheap unit of work. Test at low duration on the cheapest ty- line first.