Zhenzhen Grok3 Video

If you've been fighting Sora 2's constant 500s and capacity wall, the pack author has a recommendation you should actually listen to: ComflyGrok3VideoApi - the grok-video-3 node - is the Sora replacement they keep coming back to. The README is unusually specific about why: multi-shot generation, image-to-video with up to seven reference images, real-person mode, Chinese voiceover support, unlimited concurrency, and 10-second clips with audio. It's xAI's Grok video model, metered through the Zhenzhen relay, and it lands as a VIDEO tensor in your graph.

That seven-reference-image mode is the standout. It's a "multi-ref" workflow where you reference each image by name inside the prompt - @img1, @img2, up to @img7 - like the seedance-style multi-parameter videos. The node uploads your images and the model composes them into one scene.

The inputs that matter

• prompt - the video prompt, including the @imgN references if you're using multiple images.
• model - grok-video-3, fixed.
• ratio - 2:3, 3:2, 16:9, 9:16, 1:1.
• duration - 6, 10, or 15 seconds.
• resolution - 480P, 720P, or 1080P.
• image1–image7 - reference images (for i2v / multi-ref).
• seed, api_key, skip_error.

Outputs: video (VIDEO tensor → save node), task_id (the async handle), response, and video_url.

Install and key

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

Restart, or ComfyUI Manager → "Comfyui-zhenzhen". Key from ai.t8star.org into api_key. Token-group note: Grok video rides the veo&grok groups on the site, so make sure your key has one assigned.

The parts that bite

Two things the README is honest about. First, Grok video content is filtered - NSFW is filtered out at the model level, so this won't be your unrestricted lane, "合规使用" (use it compliantly) as the author puts it. Second, the multi-ref @imgN prompt format is mandatory: plug in seven images and write a prompt that doesn't reference them, and you've wasted a run - the model needs the @img1...@img7 markers to know which image is which. Beyond that, standard pack rules: transient 500s re-run clean, the task_id lets you find results in the site's async list, and your prompts and reference images leave the machine. For short-form Chinese-dialogue video that doesn't require a Sora badge, this is genuinely the more reliable tool.