提示词｜Grok 多图剧本重构

This is the pack's "video remix" glue node, and it's genuinely clever. Grok video models can reference up to seven images via @img1–@img7 anchors in the prompt - the hard part is authoring a prompt that actually wires those anchors to specific images. This node does it for you: it takes a video breakdown report (from the pack's Gemini video-analysis node), plus a description of the new product you want to insert, and hands back a rewritten prompt that's ready to paste into a Grok video node. The use case the README is built around is TikTok-style second-creation: take an existing video's shot-by-shot breakdown, replace the product with yours, keep the motion.

Unlike every other node in this pack, it doesn't generate pixels - it generates text, using an LLM of its own. That means it's the cheapest node here by a mile, since you're paying for a chat completion (default gpt-5.4-mini, which the tooltip calls out as best value) rather than per second of video. You still need the same sk- key, because the rewrite happens on Tikpan's relay too.

How it works

Three inputs go in:

• Gemini原片拆解报告 - paste the output of a Gemini video-analysis node here. This is your source material: what happens, shot by shot.
• 核心产品与植入场景 - describe your product and where it should appear. The default template walks you through it: product name, key selling points, the specific scene to insert into (e.g. "replace the water bottle in the protagonist's hand with my @img1 perfume, with shimmering light on the spray").
• 氛围与运镜微调 - optional extra instructions about tone, grading, camera moves.

Then the node calls the chosen text model to rewrite all of that into a Grok-3-format prompt. The clever part is the anchor system: each optional 图N_对应的主体描述 field tells the LLM what @imgN actually is ("a silver laptop", "a red sneaker"), so the generated prompt names each image correctly instead of guessing. Leave them empty and the LLM will still write anchors, but it'll be guessing what they show.

Outputs are 🎯_Grok3专属提示词 (the finished prompt, wired into your Grok video node's prompt input) and 🧠_GPT重构思考日志 (the model's reasoning, handy when the result misses).

Getting it right

The quality of the output tracks the quality of the report and the product description - garbage in, prettier garbage out. Be concrete in 核心产品与植入场景; vague templates produce vague inserts. And when the report references shots, mention which ones matter. It's a text node, so iterate freely: rerunning it costs almost nothing, which is exactly why it's worth tuning here before you spend the video credits.

Install and first run

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

Restart, or search "Tikpan" in ComfyUI Manager. Paste your key into Tikpan_API密钥, drop the report in, fill the product block, run, and read the output - it's a string, so you can eyeball it before it ever touches a video node. Note this node's 校验HTTPS证书 defaults to on (the video nodes default it off), so if you hit an SSL error here, that toggle is the first thing to check.

Common issues

It's an LLM call, so the failure modes are the boring network ones: 401/402/429 on the key and balance, and a slow or failed completion if the relay is having a moment. The interesting failure is semantic - the prompt comes back but doesn't actually use your product description. When that happens, re-read what you put in 核心产品与植入场景 and the anchor descriptions; the model can only remix what you gave it.