视频｜Grok Imagine Video 1.5 Preview 首帧生视频

A peek at xAI's next Grok Imagine video model before it's the default: grok-imagine-video-1.5-preview, first-frame only. You feed it one image, it animates that image into a short clip. That's the whole node - no text-to-video, no editing, no multi-reference. The name says "preview" and it behaves like one.

Here's the part to read twice before you build a workflow on it: the preview is not the cheap early-adopter lane. Tikpan prices it at 0.0600 per input image and then 0.4800 per second of output at 480p (0.8400 at 720p) - noticeably steeper per second than the current grok-imagine-video node's 0.3000/0.4200. You're paying a premium to see where the model is headed. For a first-frame test, that's a reasonable bet on quality. For production content at 720p, it's a pricey habit.

How it works

Identical plumbing to the pack's other video nodes, minus the mode dropdown. The model is fixed to grok-imagine-video-1.5-preview, so 首帧图 is required - that's your first frame and your image prompt in one. Your prompt in 生成指令 describes how the still should move: camera, pacing, atmosphere, and (Grok videos do this well) the sound to match. The node submits to /v1/videos/generations, polls the task, downloads the finished clip, and returns the same five outputs as the main video node:

• 🎬_视频输出 (VIDEO) - the clip as a tensor.
• 📁_本地保存路径 - where the mp4 lands.
• 🔗_视频云端直链 - hosted URL.
• 🆔_任务ID - your rescue handle on timeout.
• 📄_完整日志 - raw API response.

The inputs that matter

• 首帧图 - required, the image to animate.
• 生成指令 - describe motion, lens, rhythm, and sound. "Turn this image into a cinematic short video with natural motion..." is the shipped default, and it's a decent template.
• 分辨率 - 480p or 720p. The price gap between them is steeper here than on the standard node, so be deliberate.
• 画面比例 - 16:9, 9:16, or 1:1.
• 随机种子 - fixed by default (888888) so you can reproduce a clip; change it to get variation.

Install and first run

It's the pack-standard install - nothing model-specific to download, the model lives on Tikpan's servers:

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

Restart, or search "Tikpan" in ComfyUI Manager. Add the node, paste your sk- key from https://tikpan.com, load an image into 首帧图, run. Expect the node to poll for a while - it's a cloud video job, not a local sampler.

Common issues

Same video-task failure family as the rest of the pack: timeouts (raise 最长等待秒数 or keep the 任务ID and use the Task Fetcher to collect the clip later), money deducted with no file (完整日志 shows whether the render actually succeeded), and the odd case where the VIDEO output won't preview but the file is fine on disk - check 本地保存路径. One preview-specific gotcha: because this is a first-frame model, a bad input image is a bad video. Feed it a clean, well-lit still and describe the motion you want in the prompt, or you'll spend your premium per-second rate discovering that garbage in is garbage out.