🦭万相3.0全能视频@炮老师的小课堂

Wan 2.2 was the last Wan you could download. After July 2025 the numbered series went API-only - 2.5, 2.6, 2.7 - and the open-source community stopped tracking it. Wan 3.0 is that same line, one more step past where the crowd lost interest. So the first thing to know about this node: it is not running a model on your GPU. It is a paid API wrapper that reaches a Chinese relay service (dapaoAI, aka 妙笔AI工坊) and asks it to run Wan 3.0 for you. If you've been on 2.2 wondering where the new Wan went, this is where it lives now - behind a key and a meter.

It's one node in the big paolaoshi/ComfyUI-dapaoAPI pack from 炮老师的小课堂 ("Teacher Pao's classroom"), a Chinese ComfyUI creator who maintains his own relay. And it's in the "大炮AI主力维护" group, the channel the author says will keep getting support. English-reddit signal on this pack is basically zero as of mid-2026 - this is the reseller ecosystem the Western community doesn't talk about, so you're trusting the README more than a reputation trail. Just know what you're installing.

What it actually does

One node covers all four video entry points: text-to-video, image-to-video, first-and-last-frame video, and multimodal reference (up to 10 images, 5 audios, and - only on the wan3.0-video model - 5 reference video clips). That's genuinely convenient. Compare it to your local Wan 2.2 rig where I2V means a Kijai wrapper, a VAE, a text encoder, and half an hour of waiting; here you pick a mode, drop in media, and the work happens elsewhere.

Model choice is simple: wan3.0 (default), wan3.0-fast (cheaper, faster), and wan3.0-video (the only one that accepts reference videos). Everything supports 720P/1080P, 自适应(auto)/16:9/9:16/1:1 ratios, 2–30 second clips, and a 🔊 生成音频 toggle.

How it works

Under the hood it's a relay client, not a sampler. Reference media gets preprocessed locally first - images downscaled so the longest edge stays under 2K, videos transcoded to MP4/H.264/AAC at ≤1080p/30fps, audio to MP3 - then uploaded once and reused by content hash (asset:// references). Your prompt and asset URLs go to https://api.dapaoai.com/v1/video/generations, and the node polls until a video URL comes back. A persistent queue with an idempotency key means a re-run with the same seed within 48 hours reuses the already-submitted task instead of double-billing you. The output is a VIDEO object the adapter downloads from that URL when a save node asks for it.

The inputs that matter

• 🔑 API密钥 - leave it empty and set the DAPAO_API_KEY environment variable instead. A key typed into the field can ride along in saved workflows; the tooltip promises keys only ever go to api.dapaoai.com.
• 🎛️ 生成模式 - 文生视频 / 图生视频 / 首尾帧生视频 / 多模态参考. Image-to-video needs 🎬 首帧图; first-last-frame needs both 🎬 and 🏁 尾帧图, or it errors on purpose.
• 📝 提示词 - in multimodal mode you reference assets by order: write "图1", "视频1", "音频1" in the prompt.
• 🎲 随机种 - this one's a trap. It only controls ComfyUI caching, because the relay's Wan 3.0 adapter doesn't accept a seed server-side. Same seed + same prompt within 48h = same video, no new bill. Want a different take? Randomize the seed.

Outputs: 🎬 视频 (VIDEO - wire into a video save node), 🆔 任务ID, 📋 响应信息 (billing and run details), and 🔗 视频URL if you want to handle the file yourself.

Installing and getting a key

ComfyUI Manager → search "dapaoAPI", or:

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

Then restart ComfyUI. Dependencies are light (requests, Pillow, aiohttp, numpy) - no model files, because there's nothing local. You do need ffmpeg on PATH if you'll feed it reference videos or audio; it errors with "未找到 ffmpeg" otherwise. Then register at dapaoAI, buy a redemption code, and mint an API key. This is pay-per-call: billed on final output seconds and model, so 402 errors mean "top up the wallet."

Where people get burned

• Re-runs reuse results. Same seed within 48h is idempotent - good for your wallet, surprising if you expected a fresh roll.
• Mode/media mismatches fail loudly: reference videos only work in multimodal mode and on wan3.0-video, and reference video clips cap at 15s total.
• Don't expect seed-level reproducibility - there is no server-side seed.
• Watch the balance, and remember everything you send - prompts, reference frames - leaves your machine for a third-party server.

The whole UI is Chinese, so if you don't read it you'll be translating labels like 文生视频 (text-to-video) on the fly. Annoying, but this is the kind of pack where the real cost isn't the language barrier - it's that every render is a paid POST to someone else's server.