心宝视频生成-即梦豆包系

ByteDance's Seedance video models - the engine behind 即梦 (Jimeng), which this node's display name points at - have no open weights. You can't download them; you call them. This node is how they enter a ComfyUI graph: a paid API call to the 心宝 channel's Doubao/Seedance endpoint, with the resulting video dropped back as the pack's VIDEO output, ready for the batch video saver.

Of the pack's three video generator nodes (Sora, Veo, and this Doubao one), this is the current model's own tier: doubao-seedance-1-5-pro-251215 - the tooltip notes it's the only model supported for now, with a list expected to expand. The KB's closed-source-models doc frames the category correctly: this is a "the model isn't on your machine, at any VRAM" case. If you want Seedance video, this is one of the doors.

The inputs that matter

• prompt (multiline) - the video prompt, and Chinese is fully supported. Advanced options aren't separate knobs - they're appended to the end of this prompt as -ratio=..., -resolution=..., -generate_audio=true/false, -camera_fixed=true/false flags. That's why the ratio, 分辨率, 生成音频, and 固定镜头 inputs all describe themselves as "appended to the prompt." Set the widget, the node builds the flag.
• banana_api_key - blank inherits config.ini. The tooltip has an important wrinkle: this channel's key is not interchangeable with the old channel's - separate billing, separate key. If your Banana generation key works but video fails auth, that's why.
• seconds - 4 to 12, default 12. The current model's duration window.
• batch_size - 1 to 8 videos per run, with a 2-second stagger between requests so you don't slam the endpoint.

Then the mode toggles: 启用工作流并发 (workflow concurrency, off by default - multiple video/Banana nodes can fire concurrently, same semantics as V2), 仅提交不等待 (submit-only: writes the task to history.toml and returns immediately; progress in the bottom-right 心宝视频任务 panel), 绕过代理, and 禁用SSL验证.

For image-to-video, there are two reference inputs: first_image and last_image (first and last frame). Provide the first frame for standard image-to-video; provide both and the node builds input_reference as a JSON array of the two URLs (it uploads each to a public URL first, which is why the tooltip recommends it for faster parsing).

The three outputs

• video (VIDEO) - the single generated clip.
• videos (VIDEO, list) - the full batch when batch_size > 1.
• text (STRING) - the task's status/error message. In concurrency or submit-only mode, this is where failure information shows up instead of a hard graph error.

Wire video/videos to XinbaoBatchSaveVideo to get files on disk, and text to a preview node so you can read what happened.

Installing and gotchas

Same pack:

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

Restart ComfyUI. No model files; the pack's requests does the heavy lifting.

Two things bite people. First, the key confusion - the Doubao channel key is separate from the Gemini route's key, and the "blank inherits config.ini" fallback only helps if the config holds the right key for this channel. Second, video is where the per-call cost is genuinely painful; the KB's Seedance reception thread put it best - a monthly subscription buys a surprisingly small number of clips. Batch of 8 at 12 seconds is eight billed generations whether you like the results or not. Start at batch 1, validate the prompt, then scale.