🦁Wan3.0-官方全能视频@炮老师的小课堂

Wan 3.0 is never coming to your GPU

Wan 2.2 is the last open Wan. Alibaba shipped 2.5, 2.6 and 2.7 as cloud previews with no weights, people asked for months whether the open line was dead, and it stayed dead. Wan 3.0 continues that pattern. So every "Wan 3.0 in ComfyUI" node you'll find is a relay, and this one is no exception: it POSTs to api.dapaoai.com, and the word "官方" (official) in its name means official Alibaba model IDs (wan3.0-video-official, wan3.0-video-prime-official) behind a Chinese reseller gateway - not ComfyUI's Partner Node storefront. You still need that reseller's key and a funded wallet.

Why reach for it when Wan 2.2 runs free on your card? Audio, mainly. 🔊 生成音频 is on by default and native audio is the thing local Wan 2.2 simply doesn't do. Then length - up to 30 seconds in one generation, where local Wan clips out around 81 frames - and multi-modal reference mode, where you hand it reference images, clips and audio and describe them by position. On a 5090, 720p Wan 2.2 is roughly 25 minutes a clip without speed LoRAs.

Which is the honest tradeoff: local Wan is free, uncensored and Apache 2.0; this is metered per second of output, and your prompts and media leave the machine. For five seconds of b-roll you could have generated locally, stay local.

What the node actually does for you

It's the plumbing nobody wants to write. Four modes: 文生视频 sends the prompt alone; 图生视频 requires 🎬 首帧图 and sends it as a first_frame; 首尾帧生视频 requires both 首帧图 and 🏁 尾帧图 and sends first_frame + last_frame; 多模态参考 assembles reference_image, reference_video and reference_audio entries from the reference slots, and you address them in the prompt as 图1, 视频1, 音频1. Reference media is only read in that mode - wire a 🎞️ 参考视频1 while in any other mode and it stops with "参考视频只在多模态参考模式使用".

Everything gets transcoded locally before upload: images scaled to a 2048 long edge (JPEG quality walked from 88 down to 72 until each is under ~8MB), video re-encoded to MP4/H.264/AAC at ≤1080p and ≤30fps, audio to MP3. Caps: 10 images, 5 videos, 5 audios; images ≤20MB total, reference video ≤15 seconds, audio ≤15MB and ≤15 seconds. With a reference video in play, the generated clip is capped at 15 seconds too.

Submission runs through a local sqlite queue with Prefer: respond-async, and identical payloads reuse the already-paid task instead of being resubmitted - a real kindness when a 1080p job is sitting in a queue. The 🎬 视频 output is a small adapter whose save_to() pulls the file from the gateway's /v1/videos/<task_id>/content endpoint (following a signed redirect without your key), so it drops into Preview Video or Save Video like anything else.

Inputs worth caring about

• 🔑 API密钥 - leave it blank and set DAPAO_API_KEY instead. That's the author's own tooltip, and the code honours it. Per the tooltip the key only ever goes to api.dapaoai.com; pasting it into the widget risks it being saved inside an ordinary workflow file.
• 🎛️ 生成模式 and 🤖 模型 - validated together; both official model IDs accept reference video.
• 🧩 分辨率 (480P/720P/1080P) + 📐 视频比例 + ⏱️ 时长(秒) (2–30). 自适应 is sent as ratio: "adaptive".
• 🔊 生成音频 - default on.
• 🎲 随机种 - be real about this one. The tooltip says plainly that the 妙笔 Wan endpoint doesn't take a seed; the widget exists to control ComfyUI's cache and re-execution. Don't expect the same clip twice.
• 🚫 出错时跳过 - returns an empty video plus empty strings instead of raising. Wonderful for long batch runs, awful for debugging: the graph goes green with nothing in it. Read 📋 响应信息.
• prompt_extend is hardcoded on and watermark hardcoded off here, so there's no widget for either, unlike the image nodes in this pack.

Outputs: 🎬 视频, 🆔 任务ID, 📋 响应信息 (mode, resolution, asset counts, task ID, elapsed time, the billing note) and 🔗 视频URL. Keep the last two around; if the download misbehaves you can still fetch the clip.

Install

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

cd ComfyUI/custom_nodes && git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI
pip install -r ComfyUI-dapaoAPI/requirements.txt   # requests, Pillow, torch, numpy, zhipuai, aiohttp, nest-asyncio, scipy

No models to download - the weights are on someone else's machine. But if you want reference video or audio, install ffmpeg and ffprobe on your PATH; the transcoding shells out to them and the pack doesn't bundle them. Miss them and you get "未找到 ffprobe，无法自动优化素材" the moment the node touches a clip. You can point DAPAO_FFMPEG_PATH / DAPAO_FFPROBE_PATH at your binaries instead.

Where it breaks

Errors are mapped from HTTP status to readable causes, which is more helpful than most relays: 401 is a bad key, 402 is an empty wallet, 403 is no model permission, 413 is media or request too large, 429 is rate-limiting or queueing - and both 413 and 429 mention oversized images, so let the node's own 2K preprocessing do its job rather than feeding it 4K plates. Reference videos outside 1–15 seconds are rejected outright.

Timeouts trip people up. ⌛ 请求超时 (180s) is the HTTP request; 🔁 最大轮询秒数 (3600s, up to 7200) is how long the node waits for the upstream job. A 1080p generation that "failed" is usually polling.

And the shared caveat for this whole category: it's a metered call on a reseller gateway, and the English-language community has essentially nothing on this pack - no threads, no reception either way - so your documentation is the README and the node's own error strings. Read what it sends before you trust it with a key.