视频｜Grok Video 1.0/1.5 视频生成

One node, two Grok video models, and they behave differently enough that the dropdown is where people trip. grok-video-1.0 is the flexible one: text-to-video, single-image, or up to seven reference images fused into the scene. grok-video-1.5 is the upgrade, but it demands exactly one reference image - no text-only mode, no multi-image fusion. If you're coming from 1.0 and leave a text prompt in, 1.5 will bounce it. The node makes the contrast explicit in the model tooltip; it's the clearest single fact about this node.

It's a Tikpan reseller node like every other in this pack: metered cloud API, your key from https://tikpan.com, footage and prompts leaving your machine, the model's filter in charge of what's allowed. Durations run 6/8/10/12/15 seconds at 480p or 720p. The KB's standing advice applies - these nodes exist because Grok Video has no local weights; that's the honest reason to use one.

How it works

You pick 模型, and the node submits to Grok Video's endpoint (the pack uses /v1/video/generations), polls the task until the clip renders, downloads it to your ComfyUI output folder, and returns four outputs:

• 🎬_视频输出 (VIDEO) - the finished clip tensor.
• 📁_本地保存路径 - file path on disk.
• 🎬_视频云端直链 - hosted URL if you don't want the local copy.
• 📄_渲染日志 - API response for debugging.

The inputs that matter

• 模型 - the 1.0 vs 1.5 decision above. That's the whole game.
• 视频时长 - 6s is a cheap test, longer costs more and takes more wall-clock time.
• 分辨率 - 480p/720p.
• 画面比例 - 16:9, 9:16, or 1:1 for 1.0; 1.5 supports only 16:9 and 9:16 (the tooltip says so - pick a 1:1 and 1.5 may refuse).
• 随机种子 - fixed at 888888 by default so runs are reproducible.
• 参考图1–参考图7 - the trap to remember: 1.0 will take up to seven and use them as scene references; 1.5 must have exactly one connected.

Install and first run

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

Restart, or use ComfyUI Manager (search "Tikpan"). Paste your key, pick a model, and for 1.5 connect a single image to one of the reference ports. Keep the duration and resolution low for the first run - video jobs here sit and poll, so a short 480p clip is how you validate the whole chain before committing to a 15-second 720p job.

Common issues

The pack-wide video failures apply: task timeouts (note this node doesn't return a 任务ID output - it's simpler than the other Grok video node, so if a long job times out you're relying on the log and the cloud link to recover), balance errors (402), and the 1.5 "exactly one reference image" rule being the most common newbie bounce. If the VIDEO output won't preview, check 本地保存路径 - the file is usually there and your ComfyUI just lacks a video preview extension. And when multiple reference images are in play, spell out each image's role in the prompt ("image 1 is the person, image 2 is the outfit"), per the pack's own docs.