Zhenzhen Grok Video Fal

This is the grown-up version of the Zhenzhen Grok video node, and the reason it exists is in the schema: seven image inputs plus a reference_image_urls text box. xAI's Grok Imagine Video can do text-to-video, plain image-to-video, and reference-to-video - where you hand it several images of a subject and it tries to keep that subject consistent across the clip. Most wrappers only expose the first two. This one exposes all three, which makes it the node you reach for when you want a character or product to survive from stills to motion.

It's part of the Comfyui-zhenzhen pack by T8mars - an API reseller ("Zhenzhen") who runs a fan-priced service out of ai.t8star.org and routes these jobs through a proxy of fal.ai's queue API (hence -fal). Closed model, no local weights, per-call credits, key in the widget. The multi-image reference mode was added in a mid-2026 update precisely so people could do multi-shot consistency without a second tool.

How it works

The node submits to https://ai.t8star.org/fal/xai/grok-imagine-video/<endpoint>, where the endpoint is chosen by mode:

• auto (default) - old behavior: one image in → image-to-video; no image → text-to-video; two or more images, or any reference_image_urls, → reference-to-video.
• text_to_video, image_to_video, reference_to_video - force a specific mode.

After submission it polls every poll_interval seconds up to max_poll_attempts (default 600 × 6s = a full hour), then downloads the result and returns a real VIDEO output. Input images go one of two ways via image_way: base64 (embedded in the request) or image_url (uploaded to the proxy first, which suits larger inputs).

Inputs worth your attention

• The seven image slots (image plus image2–image7) - the more reference images, the stronger the consistency ask, but also the slower the job and the more it costs. "Too many will affect consistency," as the pack's own changelog puts it; don't treat seven as a target.
• reference_image_urls - newline- or comma-separated public URLs appended after the wired images. Useful when your source stills are already hosted.
• aspect_ratio - note the auto option exists but is image-to-video only; on text or reference modes the node silently forces 16:9, so don't rely on auto for a vertical prompt.
• duration - 1–30s, default 6. resolution - 720p or 480p, no 1080p.
• api_key - required; empty key means "API key not provided".

Outputs are video (VIDEO - feed it into Save Video/VHS), video_url (STRING), and response (raw payload for debugging).

Installing it

Manager: search Comfyui-zhenzhen and install. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

Restart, paste your API key into the widget, done. No model downloads, no VRAM pressure - the whole point of the pack is that the GPU work happens on the vendor's side.

Common issues

• Reference mode ignoring your prompt - you've probably wired multiple images but left mode in auto with a single image connected; the mode flip only happens on the counts above. Set mode explicitly if the behavior surprises you.
• 443 / timeouts - servers are overseas; Zhenzhen's README says plainly that a VPN in TUN mode is expected, and both .org and .cn mirrors are tried automatically.
• 500 errors - frequently the vendor's load, not your setup. Re-run before debugging anything.
• Token group - this model runs on the default group; if you've edited your token's groups on the site, make sure default is still attached.

One light aside: the "seed" field here is a ComfyUI cache-control nicety, not a real reproducibility seed - the API doesn't take one, so randomize it and move on.