1.Grok 提交任务 (Wujiai)

Grok video (the xAI video model, grok-video-3 in this pack's framing) is closed-weights - you can't download it, so the only door is an API. This node is that door on the Wujiai channel (无极 AI, the second reseller this pack wraps, after Grsai). It's the numbered first step in a three-node video flow: 1. submit → 2. query status → 3. download video. The numbering is right there in the node names, which tells you everything about how the author expects you to wire it.

The flow is asynchronous by nature: submit here, poll with GrokQueryTasks, pull the finished file with GrokGetNextVideo. There's no blocking wait inside this node - you get a task reference back and the job churns on the Wujiai side while your ComfyUI stays responsive. That's the right shape for video, which takes minutes, not seconds.

The inputs that matter

• prompt (STRING, multiline, default "A cute cat") - what you want the video to be. This is the one you'll actually edit.
• aspect_ratio - one of 2:3, 3:2, 1:1, 9:16, 16:9 (default 16:9). Note there's no "auto" here; pick your frame.
• duration - 6, 10, or 15 seconds. The README's headline is that this channel supports 10s/15s long video, so 6 is the cheap baseline, 15 is the max.
• images (optional IMAGE) - attach one or more reference frames for image-to-video. This is a batch input, so wire in a preview batch or a DynamicImageList output.
• api_key (optional STRING) - leave empty and the pack falls back to the WUJIAI_KEY env var or the ApiKeyManager-configured key.

Output is a single response STRING - the submit receipt with the task info you'll need for the query/download nodes. It's not a video.

Install & prerequisites

Install once for the whole pack: ComfyUI Manager (search "ComfyUI-nkxx"), or

cd ComfyUI/custom_nodes
git clone https://github.com/jieg9341-lab/ComfyUI-nkxx

then restart. First launch auto-installs requests, pandas, openpyxl, yt-dlp, opencv-python, and aiohttp; if that fails, pip install them manually. No models download - nothing runs on your GPU.

The real prerequisite is a Wujiai API key (register at wujiai.org, per the README's own link). This is a paid reseller channel: video generation is metered per call, so every submit costs credits even if the render later fails. If you get an error about a missing key, it means the pack couldn't find one anywhere - check the node field, then the WUJIAI_KEY env var.

Common issues

• Response but no video: normal. This node only submits. If you don't see your task in GrokQueryTasks afterward, the Wujiai side may have rejected it (billing, moderation, model availability) - the report text will tell you.
• Cost surprises: unlike a local sampler, a 15s render is a real line item. Don't test with 15s; test with 6.
• Trust note: the Wujiai-side node files ship obfuscated in this pack (base64+zlib exec blobs, not readable source), so the submit mechanism is only verifiable by behavior. Worth knowing before you paste a paid key into anything.