🐠Seedance2.5-SP视频（30秒）@炮老师的小课堂

There is no Seedance checkpoint to download. ByteDance's whole video line is API-only - the two-track posture the company is famous for: components go out Apache 2.0, products stay behind an endpoint. So you're not loading a model here, you're wiring in an HTTP client with your key in it. This node is one of those, pointed at a Chinese reseller gateway (https://api.dapaoai.com, "妙笔AI工坊") rather than ByteDance's own API.

Why reach for it: a fixed 30-second clip, text-to-video or image-to-video with up to nine reference images, no VRAM, and the result lands back in your graph as a video you can save or feed onward. For boards or a single continuous shot at that length, the local alternative is a 5-second Wan or LTX clip and a prayer.

What it actually sends

The node submits POST /v1/videos with model: seedance-2.5-sp, your prompt, seconds: "30", the ratio, and the resolution lowercased. Each reference image becomes an in-memory PNG, scaled so its longest edge is 2048px (Lanczos), uploaded to the gateway and passed as an asset://file-... reference.

Two details worth knowing. Unlike the Seedance 2.0 all-round node in the same pack, this variant does not register those uploads in the asset library - that's the "不登记素材库" in its description, and it means the references are throwaway. And images are capped at 10MB and resized before upload, so pre-upscaling references is wasted effort.

It then polls GET /v1/videos/{task_id} every 10–15 seconds until the job finishes. The 🎬 视频 output is a lightweight adapter object with get_dimensions() and save_to(), which streams the finished file from the gateway's /content endpoint with your key. That duck-typing is why it plugs into SaveVideo, VHS and most other video nodes.

Inputs you actually touch

• 🔑 API密钥 - required. The author's tooltip is honest: it's only ever sent to api.dapaoai.com and isn't written to a config file. It does live in the node widget, so don't hand out workflow JSONs with it filled in.
• 🎛️ 生成模式 - 自动识别 / 文生视频 / 图生视频. Auto goes image-to-video as soon as any reference is wired, which is usually what you want. Leave it manual and the node rejects a mismatch instead of guessing.
• 📝 提示词 - up to 3000 characters. Longer than that and it errors before spending anything, which is the right order.
• 🧩 分辨率 - 480P to 4K, default 720P. The author's own tooltip says "协议允许，尚未逐档实测": allowed by the protocol, not verified tier by tier. The code forwards it and computes an expected size locally for the adapter, so treat 2K/4K as unproven until you've paid for one.
• ⏱️ 时长(秒) - one option, 30. Locked. Any other value throws.
• 🎲 随机种 - cache control only; it is not sent to the API. Its only job is to make ComfyUI re-run the node. You cannot recall a specific take with it, so change the prompt to explore.
• 🖼️ 参考图1–9 - nine slots, but the cap is nine images combined, so a batch of two counts as two.

Outputs: 🎬 视频 (wire to a save node), 🆔 任务ID (keep it - it's your support ticket and your only proof of what you paid for), 📋 响应信息 (a readable summary that points you at your gateway billing instead of quoting a price), 🔗 视频URL.

Installing

ComfyUI Manager → search dapaoAPI → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git
cd ComfyUI-dapaoAPI
pip install -r requirements.txt

The requirements file is heavier than an HTTP node deserves - torch, numpy, zhipuai, aiohttp, nest-asyncio, scipy - because the pack also ships Gemini, GLM and RunningHub nodes. Then get a key: register on the author's referral link, buy a redemption code into your wallet, request a key in the default group, paste it in. The README is blunt that only the 妙笔 and RH channels in this pack are still maintained - ignore the older provider nodes.

Where people get burned

Mode/input conflicts. "文生视频不能同时接入参考图" and "图生视频至少需要一张参考图" are validation, not bugs - pick 自动识别 and move on.

Stale widgets from the sibling node. This variant deleted 首帧图, 尾帧图, 生成音频 and the video/audio reference slots the Seedance 2.0 node has. A graph still carrying them fails with "Global SP不支持这些输入：…", which beats silently dropping your first frame.

Long queues blow the poll budget. Max polling defaults to 1800s and 轮询间隔 is clamped to 10–15s. Raise 最大轮询秒数 (up to 7200) at busy times; lower 请求超时 if you'd rather fail fast on the upload.

Don't resubmit after a timeout. If the node says it didn't get a valid task ID, it deliberately refuses to send the job again. You're billed per call - keep the ID and query it.

Cost and filtering. This is a reseller in front of ByteDance's model: per-second metering, and the "looser moderation" resellers advertise is marketing - your prompt still hits Seedance, which does its own refusing. As for the key: read the source before you paste one - this pack is MIT and plain Python, and it talks to exactly one host.