视频｜Grok Imagine Video 音视频生成/编辑

Grok Imagine Video is xAI's closed video model, and this node is the Swiss-army version of it inside the Tikpan pack: one node that does text-to-video, first-frame (image-to-video), reference-image video, and video-to-video editing depending on a single dropdown. If you only install one video node from this pack, it's a decent bet - you get four workflows without learning four nodes.

The catch is the price model, which is worth reading before you run anything long. Tikpan bills per second of output video: 0.3000 per second at 480p, 0.4200 at 720p, plus 0.0120 per input image and 0.0600 per second of input video when you're editing. A 10-second 720p clip lands around 4.2 units plus inputs. This is a metered cloud API, same as every node in this pack - your footage leaves the machine, and the model's content rules apply.

How it works

Pick 生成模式 and the node does the rest: it builds a payload for Tikpan's /v1/videos/generations endpoint (or /v1/videos/edits when mode is video_edit), submits the task, then polls the task endpoint every few seconds until the clip is ready. On success it downloads the file, saves it under your ComfyUI output directory, and hands back five outputs:

• 🎬_视频输出 (VIDEO) - the tensor you can preview or feed downstream.
• 📁_本地保存路径 - where the mp4 actually lives.
• 🔗_视频云端直链 - the hosted URL, useful if you'd rather skip the download.
• 🆔_任务ID - needed if the job times out and you want to fetch it later.
• 📄_完整日志 - the raw response, your debugging friend.

The inputs that matter

• 生成模式 - text_to_video (prompt only), first_frame (needs 首帧图), reference_image (needs 参考图), or video_edit (needs a source video).
• 分辨率 - 480p vs 720p. The tooltips are blunt about it: 720p bills higher per second.
• 画面比例 - 16:9, 9:16, or 1:1.
• 随机种子 - default 888888; same seed + same prompt reproduces the clip.
• 最长等待秒数 (default 1200) and 查询间隔秒数 (default 8) - how long the node polls before giving up. Long or 720p jobs sometimes need a higher ceiling.

For video_edit, you can feed a local VIDEO tensor, a 本地视频路径 (mp4/mov/webm), or a 视频URL. The docs flag a real limit here: big local files hit the inline upload cap - under about 18MB inline is fine, anything bigger should go through a URL.

Install and first run

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

Restart (or ComfyUI Manager → search "Tikpan"), paste your sk- key from https://tikpan.com, set a mode, write the prompt. Video takes a while - expect the node to sit polling for a minute or more, with a progress bar tracking it.

Common issues

Video is where timeouts and partial failures actually bite, and this pack's docs are refreshingly specific about them. If the task times out, you still have the 任务ID - grab it and use the pack's Task Fetcher node to poll for the finished clip instead of paying for a re-run. If money was deducted but no file arrived, the log's task_id and cloud_url tell you whether the clip actually rendered. 跳过错误 returns empty outputs instead of killing your workflow. And if 🎬_视频输出 won't preview, the file is usually still on disk - check 本地保存路径, since a missing video-preview extension is a local ComfyUI thing, not a generation failure.