AI Subway Outfit-Change Video Reverse-Engineering Prompt

Here's the thing.

This kind of AI subway outfit-change video has been everywhere lately. @GeekCatX (知识猫AI实验室) puts it plainly, these clips are pulling hundreds of thousands of views each, almost effortlessly. And in his words, at his place you can replicate the whole thing the easy way.

His setup, a reverse-engineering Skill optimized with GPT-6 Astra that outputs everything in one pass, running on MiniMax H3's image-to-video mode. He adds, with a bit of self-mockery, that he's broke and can't afford Seedance 2.5, so H3 it is. The result, he insists, holds up just fine.

The clip itself is the format in miniature. A young woman sits on a subway seat scrolling her phone, and in the foreground another phone screen shows a shopping app, a lace bodysuit and stockings on the product page. The video runs the transformation from that in-app pick to a new outfit on the same person, in the same seat, under the same train lighting.

The full prompt lives in the replies of the original thread, x.com/GeekCatX/status/2096513305930297353

There's backstory here too. The quoted post explains it. The big video reverse-engineering engine he used to run had to shut its demo down, the token costs got too heavy. Instead of waiting for a reset, he built a lighter version himself, a MiniMax H3 video reverse-prompting template.

How it works, you feed it a reference video and it reverse-engineers an H3-ready prompt from it. It supports both text-to-video and image-to-video, and once the reverse pass is done you can polish the prompt against the official H3-Prompt Writing guide for an extra bump in quality.

That template sits behind a paid beta for his community, 299 yuan for permanent beta access plus his ongoing low-cost AI research tricks. The subway prompt itself, though, he dropped free in the replies, which is what this entry covers.

At the time of writing, the post sits at 307 likes, 278 bookmarks and 44.4K views, and the quoted post has run past 214K views.