🎬豆包 Seedance 参考生成视频

This is the most powerful node in the YV919/comfyui pack, and also the one with the most footguns. 参考生成视频 folds three of the pack's six features into one node - video editing (rewrite a clip from a reference), video extension (continue an existing video), and multi-modal reference (mix images, video, and audio into a new generation). It's the pack's answer to the official ByteDance "Seedance reference" node, minus the parts the DMXAPI endpoint doesn't expose (auto_upscale, web search, last-frame PNG).

The headline trick is the interface: the sockets grow as you connect them. Wire one image and image_2 appears; connect a video and video_2 pops up. It's the ComfyUI V3 autogrow pattern, and it's genuinely nice once you stop being startled by inputs appearing out of thin air.

How it works

Like every node here, it POSTs a Seedance 2.0 request to the DMXAPI endpoint and hands back a task_id for the 🎬豆包 Seedance 获取 node to poll. What's different is the media plumbing, and it has rules:

• 图 (image_1..9, IMAGE) - wire Load Image nodes. Up to 9.
• 视频 (video_1..3, VIDEO) - wire a Load Video node. This is where the privacy footgun lives: the DMXAPI endpoint only accepts public URLs for videos, so the node anonymously uploads your local video to litterbox.catbox.moe (a temporary file host, link auto-expires after 72h) and sends that URL onward. Fine for anything you wouldn't mind public, actively bad for private material.
• 视频链接 (video_url_1..3, STRING) - the safe alternative: paste a public URL or asset:// ID directly. No upload, nothing touches a third-party host. This is also how you do video extension - feed the fetch node's 视频链接 output into video_url_1 and the next job continues the previous clip.
• 音频 (audio_1..3, AUDIO) - local audio auto-converts to WAV base64 (the official API only accepts wav/mp3; FLAC errors). Each clip must be 2–15 seconds, ≤3 clips, 15s total.
• 素材 (asset_1..9, STRING) - material IDs; bare IDs get asset:// prefixed automatically. One honest limitation: DMXAPI can't resolve what type an asset is, so only image-type assets actually work - video/audio assets get submitted as images and may fail. The README flags this as unconfirmed territory.

The constraints that trip everyone: you need at least one image, one video, or one asset (audio alone won't do), images+assets cap at 9, videos (both kinds) cap at 3, audio at 3. In the prompt, reference inputs by bracket index - [图1], [视频1] - in order (images → assets → videos → audio).

Inputs that matter

• 提示词 - default: "主体按提示词运动；可参考[视频1]的运镜". Reference your materials by bracket index and describe the edit, not the scene from scratch.
• 视频 vs 视频链接 - private clip? Use video_url. Public or throwaway footage? video_1 is convenient. The tooltip on video_1 is explicit: private videos, do NOT use this input.
• 生成音频 - here it's cosmetic: DMXAPI documents that this flag has no effect in reference mode, and the video always carries an audio track. Leave it.
• 分辨率 / 时长(秒) / 宽高比 / 种子 / 水印 - the pack's standard set; 480p and short duration for your first test, because every run is a billed API call.

Outputs

任务ID and 检索模型, both into 🎬豆包 Seedance 获取, which polls and drops the finished MP4 into output/doubao/.

Installing

Manager → search comfyui-doubao-seedance (or 豆包 Seedance), or:

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

Restart ComfyUI completely afterward. No extra dependencies and no model downloads - it's a cloud API pack - but it does need a recent ComfyUI for the V3 autogrow API. The pack ships a pre-wired 豆包参考生成视频工作流.json example.

Common gotchas

• 参考生成视频至少需要 1 张图 / 1 段视频 / 1 个素材ID - you connected only audio, or nothing. Add one image/video/asset.
• 参考视频只支持 URL 或素材ID…不支持本地文件 - you typed a local path into a video_url field. That field only takes URLs/asset IDs; local video goes into the video_1 VIDEO socket (which uploads it) or is hosted somewhere yourself.
• "无法拉取视频 / 生成失败" from the API - the free file-host domain got blocked on the provider's side. Switch to video_url_1 with your own public link.
• Audio "格式错误" - an old pack build encoded FLAC, which Seedance rejects; update the pack and fully restart. Also check your clip is within 2–15s.
• Over the limit - "参考图 + 素材最多 9" / "参考视频最多 3" / "参考音频最多 3" are hard caps, not suggestions.
• Cost - multi-modal reference bills like any other job here; validate the edit at 480p before a long 1080p reference run.