🦁Minimax-H3全能视频@炮老师的小课堂

MiniMax H3 is the 33B omni-modal video model everyone got excited about in August 2026 - picture and stereo audio in one pass, not a glued-on audio step. You are not running those weights here. This node is an HTTP client with a paid relay behind it, and it'll hand you an H3 clip on a laptop with no GPU - either the best thing about it or the reason to skip it, depending on how you feel about paying by the second.

The behaviour is specific, though: up to nine images, three reference videos, three audio files, mode inferred from whatever you connect, and - the part I'd install it for - the MP4 saved to disk before the node returns, so a closed tab doesn't cost you the render.

What it's actually doing

Nothing local - that's the whole point and the whole cost. An API node is arbitrary Python that holds your key and calls the network by design, so the risk it carries is exactly the convenience it sells.

In order: images get downsampled to a 2048px long edge, video transcoded to MP4/H.264/AAC at ≤1080 and 30fps, audio to MP3; everything uploads through a content-addressed asset endpoint, so a file you've sent before isn't re-sent; then one POST to /v1/videos via a persistent queue (Prefer: respond-async plus an idempotency key), polling until completed, streaming into ComfyUI/output/dapao_h3/, and checking for a real video stream before the atomic rename. The paid POST is never retried - a retry is a double charge - and a receipt (request JSON, IDs, local path) lands in ComfyUI/user/dapaoAPI/h3_tasks/, keyed by a hash of your key rather than the key itself.

The inputs worth setting

🔑 API密钥 can stay blank if you set DAPAO_API_KEY in the environment - do that if you share workflows, because the widget value serialises into the graph. 🤖 模型 has one choice, minimax-h3-miaobi: one relay channel, not the official MiniMax endpoint, so nothing to flip. 📝 提示词 wants a shot description.

Then 🧩 分辨率 - 720P or 1080P, no 2K here - 📐 视频比例 (16:9 or 9:16, the whole list), ⏱️ 时长(秒), a whole number from 3 to 15, and 🔊 生成音频, on by default.

Which of 🖼️ 参考图1…9, 🎞️ 参考视频1–3, 🎵 参考音频1–3 you connect decides the mode: nothing is text-to-video, one image is image-to-video, and several images plus a reference clip and a music bed is the "all-round" case it's named for. Audio alone is rejected; it needs an image or video to hang off.

🎲 随机种 is not a model seed - its tooltip says cache and task recovery, and it never reaches H3. With the generate control fixed, re-queueing an unchanged graph returns the cache for free: a spend guard, and why people think it "won't make a new one."

🆔 继续任务ID does something genuinely different: paste an original job_… queue ID or a video task ID and it skips the prompt, the upload and the generation POST - polling and saving only. That's your move after a timeout or failed download. The knobs beside it (🔁 最大轮询秒数, ⏱️ 轮询间隔, ⌛ 请求超时) are leave-alone unless long 1080p jobs keep dying.

Outputs

🎬 视频 is a real ComfyUI VIDEO for a save or preview node, though the file already exists. 🆔 任务ID is what you paste back into 继续任务ID later - keep it. 📋 响应信息 is a text receipt; 🔗 视频URL is the upstream link.

Installing it

ComfyUI Manager, search dapaoAPI, install, restart, hard-refresh. Or manually:

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

Requirements are light (requests, Pillow, torch, numpy, aiohttp). What's missing and does matter: ffmpeg and ffprobe, which preprocessing and the download check shell out to. Not on PATH? Set DAPAO_FFMPEG_PATH / DAPAO_FFPROBE_PATH.

Where people get burned

The limits are hard rejections, not silent clipping: over twelve items total, more than nine images, audio with nothing visual attached, a clip under two seconds, a class over fifteen seconds. All raise before anything uploads - correct behaviour that reads as "the node won't run" if you batched nine frames onto one IMAGE wire, because batches count frame by frame. An image over 10 MiB after the 2048px pass is refused too; crop it.

Cost is the other one. The node reads the public price table and shows roughly ¥0.075/second at 720P and ¥0.135/second at 1080P for the standard group - estimates only; the final bill is the relay's business. Fifteen seconds at 1080p isn't cheap, so write the prompt before you re-queue.

Be clear-eyed about the pack, too: the author is a Chinese AI teacher whose README is a stream of promotions, and the node ships a register-your-key button with his referral code in it. The code is honest; the risk is depending on one person's relay staying online and solvent. For a vetted path, the official ComfyUI API Nodes list MiniMax as well: pricier, minus the resume-by-ID trick.