Grok Imagine Video Edit (xAI)

Why you'd reach for it

The headline move in video this year was editing, not generating - same reason Kontext reshaped image work: reshooting a whole clip to change one detail is the expensive way to do it. This node takes an existing MP4 and a sentence, and gives you the clip back with that one thing changed. The rest is preserved. "Give the woman a silver necklace." Done.

The workflow that makes it click is chained: make a clip with Grok Imagine Video, dislike the wardrobe, and instead of paying for a fresh generation you pipe video_path straight into this node and edit it. It's a second pass, not a second shoot.

Be clear about what it is though: this is prompt-level editing, not regional V2V. There's no mask input, no per-frame control. If you want to repaint one region across every frame of a clip, that's a local workflow, and people ask for exactly that and don't find it here.

Inputs

video_path is the required one - either the video_path output of a Grok video node or a path you type in. The hard limit is 8.7 seconds, max, and the accepted codecs are H.264, H.265 and AV1. Over-length files fail at the API, so check before you queue:

ffprobe -v error -show_entries format=duration -of default=nw=1:nk=1 myclip.mp4

There's also video_url, an optional alternative for a public https URL of an MP4, used only when video_path is empty. Handy when the file already lives somewhere on the internet and you'd rather not re-upload it.

prompt describes the change and nothing else - don't re-describe the whole scene. model defaults to grok-imagine-video (1.0), which is the one xAI documents for editing; the 1.5 ID is in the dropdown too, but 1.0 is what the author's tooltip points you at. seed is the usual cache-breaker: not sent, not reproducible, and a fixed value with unchanged settings means a cached run that costs nothing.

The output keeps the input's duration and aspect ratio, capped at 720p - so an 8-second edit comes back as an 8-second 720p clip, no re-framing to argue about.

Outputs, uploads and the rest

The output set is identical to the other video nodes: frames, fps, audio, video (native VIDEO object), video_path, prompt, info. video_path goes to Grok Video Save; frames feeds anything IMAGE-shaped - an upscaler, a frame grab, VHS-style nodes; audio is a real AUDIO dict, and info reports the model, duration, the moderation flag and the billed cost.

One mechanical detail worth knowing: your local MP4 is uploaded inline as base64, not handed over as a URL. That means upload time scales with file size, which is another reason the 8.7-second cap is a blessing - trim first, edit second. extract_frames and extract_audio default on; both are media decoding done locally, so switch off whichever you don't need and save yourself the RAM. show_preview plays the finished clip with sound inside the node. The download lands in your temp folder (%TEMP%\grok_imagine) unless you set download_folder, and the save node moves it out of there.

Install

ComfyUI Manager, search Grok Imagine. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Slartibart23/comfyui-grok-imagine-video
..\..\python_embeded\python.exe -m pip install -r comfyui-grok-imagine-video\requirements.txt

Restart ComfyUI and press Ctrl+F5 once in the browser - that's what loads the in-node preview script. The dependencies are light (requests, numpy, Pillow, imageio, imageio-ffmpeg), with ffmpeg coming from imageio-ffmpeg unless you already have one on PATH. The API key comes from the api_key widget, then XAI_API_KEY, then grok_api_key.txt in the ComfyUI root, the portable root or the pack folder. Use the file or the env var; the widget value travels inside any workflow you share.

Where people get burned

The 8.7-second cap is the one that wastes your time - an over-long input comes back as an API error, not a cropped edit. Impatient polling: poll_timeout_seconds is 900 by default, and while edits are faster than a 15-second 1080p generation they still take minutes at times; raise it rather than re-queuing and paying twice. Editing something the API won't touch - filters apply here exactly as they do on generation, and info shows the respect_moderation flag when a result comes back oddly unchanged. And if you're on a 2.x install that predates these nodes, the whole video/save group needs deleting and re-adding once to pick up the current widgets.