🎬豆包 Seedance 文生视频

Here's the thing: you will never run Seedance 2.0 on your own machine. ByteDance keeps its flagship video model API-only - no weights, ever - which puts it in the same "call it or don't" bucket as Sora and Veo. So if you want Seedance from inside ComfyUI, you have exactly two doors: Comfy's official Partner Nodes (prepaid credits, requires logging into your Comfy account), or a reseller pack like this one, which routes through DMXAPI with a plain sk- key. This node is the entry point: prompt in, a task ID out, no GPU involved at all.

That last part is the real draw. This works on a laptop with integrated graphics, a Mac, a cloud VM with no video card - anywhere ComfyUI runs. The trade is that nothing happens locally. Your prompt leaves the machine, gets moderated by ByteDance's filter, renders on their hardware, and you pay per call. That's the category trade in one sentence: an API node is the right tool for a model you cannot run, and the wrong default for one you can.

How it works

The 文生视频 node builds a payload - model name, your prompt, ratio, resolution, duration, audio flag, seed, watermark - and POSTs it to {base_url}/v1/responses with your DMXAPI key as a Bearer token. The response is a task_id (a cgt-... string) plus the retrieval model seedance-2-0-get. That's the whole job of this node: it submits, it doesn't render.

So the workflow is a two-step pipe, and the second step is mandatory:

🎬豆包 Seedance 文生视频 ── task_id ──▶ 🎬豆包 Seedance 获取
                      ── get_model ─▶

Drag both outputs to the fetch node's matching inputs, and the fetch node polls, downloads, and previews the MP4.

The inputs you'll actually set

• 提示词 - a plain description. The default is a Chinese daisy-field example; the model understands Chinese and English both. Describe motion and camera moves ("镜头拉近定格在…", "360度环绕运镜") - Seedance does well with explicit direction.
• 密钥(api_key) - your DMXAPI token (sk-xxxxx) from dmxapi.cn, with balance on it. Same key goes in the fetch node.
• 宽高比 - adaptive, 16:9, 9:16, 1:1, 4:3, 3:4, 21:9. Pick before you spend; this sets the frame.
• 分辨率 - 480p / 720p / 1080p. Start at 480p. Video is billed per job, and a 480p 5-second test costs a fraction of a 1080p run. Confirm the style works, then scale up.
• 时长(秒) - 4 to 15. The default 5 is a good starting point; longer is where cost climbs.
• 生成音频 - on by default; Seedance 2.0 can produce synced audio (mono voice/SFX/BGM) in the same pass.
• 种子 - -1 for random; a fixed value reproduces similar results. Set 生成后控制 to randomize if you want a fresh video every run.
• 水印 - off by default; flip it on if you want the watermark burned in.

The remaining fields - 模型 (defaults to doubao-seedance-2-0-260128, leave it) and 接口地址 (defaults to https://www.dmxapi.cn) - you can ignore.

Outputs

任务ID and 检索模型 - both feed the fetch node. Nothing else. This node is deliberately thin; the reference/multi-modal nodes in the same pack cover the cases where you have images or video to attach.

Installing

ComfyUI Manager → search comfyui-doubao-seedance (or 豆包 Seedance) → install. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/YV919/comfyui

Then quit ComfyUI entirely and relaunch - refreshing the page won't load new node code. No extra dependencies (the pack uses requests/Pillow/numpy/av, all already in the ComfyUI env), but it needs a recent ComfyUI for the V3 node API. The README also ships four ready-wired example workflows, including one for exactly this text-to-video path.

Common gotchas

• 请填写 DMXAPI 的 api_key - key missing on one of the two nodes. Both need it.
• "Waiting forever, nothing in output" - check the fetch node's 最大等待(秒). Slow renders can exceed the default 300; bump to 600.
• Cost surprise - people underestimate how fast per-call video billing adds up; the community reaction to Seedance-in-ComfyUI pricing was famously "extremely expensive." Test cheap first, budget before a long batch.
• If you accidentally grab the 首帧生视频 node for a pure text job - it will error asking for a first-frame image. This 文生视频 node is the one with no image inputs; that's the correct pick for text-only.