CrePal Becomes First "Fat Cat" Video Agent

"It even does regional trash talk. Cè nà."

In my article on MiniMax's Pangmao Universe, I made what I believe was a genuinely constructive proposal:

"MiniMax should develop a Pangmao Agent, complete the Pangmao video workflow loop as soon as possible, and let Pangmao creators achieve McDonald's freedom."

I didn't expect that just one month later, a young man from Northeast China, @jiaming, would become the first to develop the world's inaugural Pangmao video Agent — CrePal.

CrePal's main selling point is one-stop AI video production.

It has a master Agent that handles task scheduling, with four sub-Agents underneath: script planning, image generation, video generation, and video editing. The sub-Agents integrate the major image, audio, and video generation models, including Midjourney, GPT-Image, Google Imagen, Suno, Hailuo AI, Keling AI, Dreamina, Pixverse, and Veo 3.

(At the end of this article, @jiaming gives his unfiltered takes on the strengths and weaknesses of each model.)

My first reaction upon opening the CrePal website was: isn't this exactly the Pangmao Agent I envisioned?

Here's a Pangmao video I generated with CrePal.

The soundtrack is surprisingly artsy, the animation more creative than what typical Pangmao creators produce — it gives off a Sanmao the Vagrant vibe.

A young man from northern Jiangsu, Pangmao has finally scraped together enough money to visit Shanghai. The Bund is bustling with crowds; Pangmao's stomach growls with hunger. He's never been to Starbucks, can't afford artisanal Italian gelato either. Spotting a familiar Mixue Ice Cream & Tea, he leans against the Bund railing, gazes out at the glittering Lujiazui skyline, and takes a sip of his large fully-sugared, lightly-iced Triple Boss milk tea for just 9 yuan — before he can even let out a satisfied sigh.

The Oriental Pearl Tower goes red-hot, charging up. Rings of energy halos rise one after another, and a ruby laser beam infused with the orthodox chill power of Jufulong precisely blasts the milk tea from Pangmao's hands. Poor Pangmao is sent flying by the shockwave, all the way up to the atmosphere to gaze upon the Milky Way 😭

The above is my imagination running wild. In reality, I only fed CrePal a single paragraph of text, along with an official Pangmao video from MiniMax: https://v.douyin.com/aFi7mOJgJHw/.

Though I was quite satisfied with this Pangmao video, I have to be honest. CrePal has only been out for about ten days, and its engineering isn't fully stable yet. This video didn't succeed on the first try — the developer sent it to me manually (@jiaming).

Jiaming was quite pleased with the Pangmao concept himself, and generated a few more Pangmao videos. In one of them, Pangmao even says "cè nà" on his own.

Cè nà, how does CrePal know Shanghainese? It turns out a fellow Northeastern Chinese, long subjected to regional stereotyping, created the world's first video Agent that independently learned regional trash talk 😭

Over the course of a week, I ran 11 tasks on CrePal. Nine of them successfully generated videos on the first try.

The two failures: one was this Pangmao-drinking-milk-tea task, and the other was when I input a live-action video hoping to adjust a character's skin tone. But CrePal only supports generating content from scratch, not modifying existing footage.

Jiaming explained that CrePal's positioning is to let ordinary users easily create high-quality AI videos. His target user is very specific: self-media creators who want to make AI videos. Pangmao creators, for instance, or those making ASMR videos with Veo 3.

You briefly describe your idea, and CrePal writes a script based on it, then draws storyboard images from the script, generates video clips from those storyboards, produces voiceover and soundtrack simultaneously, and finally edits everything into a finished piece.

Users no longer need to switch between multiple model platforms or learn various prompt techniques. The Agent handles the workflow scheduling.

This genuinely fulfills my dream of becoming a Pangmao creator. I made 7 Pangmao videos with CrePal.

The hottest meme in AI video lately has been Qin Shi Huang riding a polar bear. I decided to innovate boldly: I had Qin Shi Huang ride Pangmao instead.

The "Hua Li Xian" background music is an MP3 I uploaded to CrePal myself. CrePal's voiceover and soundtrack capabilities are fairly basic. It can synthesize voiceover based on the tone of a video you input, but this custom voice can only recite, not sing. CrePal can only generate background music through Suno, or use music you upload yourself. For example, I wanted Pangmao to sing to the melody I uploaded, with lyrics like "Pangmao rides Qin Shi Huang, Qin Shi Huang also wants to ride Pangmao." But that wasn't possible — CrePal could only generate a new BGM and have Pangmao recite the lyrics in a sarcastic childlike voice. I also encountered a character consistency issue. When generating a video of Northeastern internet celebrity Yujie riding Pangmao, I uploaded a photo of Yujie for reference, but the Yujie CrePal generated looked nothing like her, and her appearance was inconsistent across two video clips.

To solve character consistency, CrePal uses an approach where it first generates a unified character image, then uses that image as the first frame to generate video clips. But it still occasionally makes mistakes.

Despite these minor issues, CrePal has already fulfilled my dream of becoming a Pangmao creator.

CrePal-generated Pangmao videos far surpass professional Pangmao creators in creativity and visual richness. The title of world's first Pangmao video Agent is well-deserved 👍

The small problems are easily fixable too. For instance, CrePal can't sing lyrics to a melody. But MiniMax's Hailuo AI can do exactly that — Pangmao creators have been using Hailuo AI to generate songs like "Qin Shi Huang Rides a Polar Bear."

I solemnly recommend that MiniMax collaborate with CrePal as soon as possible, combining Hailuo AI's voice generation capabilities with CrePal's video production scheduling to jointly empower the world's first Pangmao video Agent.

Of course, CrePal can do serious videos too.

I specified that it call Veo 3 to generate an ASMR video of Pangmao cutting a hamburger. The one-shot quality was decent. Though the post-cutting physics don't quite make sense, that's a model limitation that can only be solved through multiple generations and cherry-picking.

After a week of testing, I generated 10 videos with CrePal.

My first impression: video generation is expensive. CrePal gives 500 credits upon registration. I paid $19 for a monthly membership with 2,000 credits, and jiaming gave me another 3,000 credits. After 10 short videos, I basically used them all up.

The cost to generate one video on CrePal is roughly $2-3. I asked jiaming whether the $19 membership pricing is profitable.

My fellow Northeastern Chinese was refreshingly honest: it's been live for less than a month, too early to tell if it's profitable. (Compare this to those who dare calculate ARR after just ten days online 🤓)

Finally, I had jiaming give his unfiltered takes on the major image and video generation models — some know-how earned through entrepreneurial trial and error.

Midjourney: Excellent control over realistic details, especially suited for photorealistic portraits or dreamcore-style images. Face consistency is also quite good — if you want to make an image of Liu Yifei, it can preserve her face with reasonable fidelity.

GPT-Image: Currently the best model for multi-image reference-based generation. It can effectively blend character images with scenes while maintaining consistency for both.

Google Imagen: Good value for money. Decent results for scene generation.

Veo 3: Very strong, stronger than all other models. If budget allows, everyone should just use Veo 3. But it is genuinely expensive — Google's official pricing is $6 for 8 seconds.

Hailuo AI: Very distinctive strengths. Its motion shots follow physics well. For footage of people running, fighting, etc., its stability is strong. (Xianyu's note: Hailuo AI is the pioneer of Pangmao videos 👍)

Keling AI: Solid all-around, no particular standout specialty, but everything works well. Good stability, low defect rate. Not too expensive, affordable. If you want reliable output without constant tweaking, choose Keling AI — it usually works on the first try.

Dreamina: Biggest highlights are cheap and fast. Beyond the low price, it also handles stylized content well, making ink-wash style images move dynamically. But it has obvious issues with physics bugs — like characters flying into the sky while running.

Pixverse: A budget alternative to Keling AI 2.1.

(Images in this article generated by CrePal, writing assisted by Gemini 2.5 Pro.)