Making an AI Morph Video: 80 RMB Failed, 35 RMB Worked

One day I was scrolling Xiaohongshu and Bilibili. Most of the feed was AI, Codex, and DeepSeek; the recommendation system had stopped pretending it did not know me. In the middle was a game-skin transformation video. The character stood in place while the outfit, hair, and weapon changed continuously. It looked genuinely good.

My first thought was simple: I want to make one too.

I picked Gwen from League of Legends and chose six skins to line up. The first round used Alibaba Cloud's HappyHorse. I connected the API, expanded images, supplied first and last frames, ran image-to-video, and eventually received an mp4.

It was an mp4 that was difficult to praise. I stared at it for a while trying to find the good parts. Gwen barely moved for several seconds, then changed outfit near the end; the source art was narrow, so the expanded sides displayed every bad guess in black.

DeepSeek helped draft the first prompt, including the instruction that “the final image settles completely into the second frame.” It sounded precise. The model read it literally: if the important part is “final,” stay still first and hand in the assignment at the end.

I wanted to try another couple of rounds. The bill calmed me down. More than 80 RMB had produced video, but not a result I wanted to publish. I was no longer making an AI video; I was paying to watch my parameters fail.

The 80 RMB did illuminate two useful mistakes

The first result was unusable, but it did separate the problems.

The first was the source material. I had used 308×560 loading-screen art with both sides already cropped. Asking a model to make that widescreen means asking it to invent the missing scene. A bad result should not have been surprising.

For the second attempt I skipped expansion and went to Riot's Data Dragon for complete 1215×717 splash art. When the official wide image exists, there is little reason to pay a model to guess what Riot painted outside the crop.

The other problem was the prompt. The first and last frames already tell the model where it must end. The useful instructions concern the twenty-something seconds in between: keep moving, make visible progress at every moment, and do not jump at the end.

MiniMax had just released H3 with 2K first/last-frame video at a lower price. I decided to pay tuition again, but this time I read the lesson first.

Six images were only the entrance; the dirty work came afterward

The six skins were not sent to the model at once. I arranged them into five adjacent first/last-frame pairs: base Gwen to Space Groove, Space Groove to Cafe Cuties, all the way to the frog raincoat. Each clip was nominally five seconds at 2K. The last frame of one clip was the first frame of the next, so they could be joined directly.

The picture was only part of it. Five clips had to become one video; five voice lines had to be found inside a long Chinese voice compilation and placed on the right moments; the theme music needed to step back whenever dialogue began.

After the expensive visual step finished, I assumed the rest would be easy. Finding voice lines was more tedious than the morphs. I let speech recognition narrow down the long recording, then cut uncertain sections into smaller pieces, listened again, and trimmed the silence around each line.

The five lines are scattered across the 26-second result and do not follow their order in the source recording. Once they were placed, the background music had to dip under each line and return afterward. None of this was difficult alone. Miss one piece and the result still felt like test clips stitched together.

All five morphs eventually followed the same rule: stop emphasising the final frame and put “constant motion, visible change at every moment, no sudden jump” first. The keyframes already tell the model where to end. The prompt's real job was to stop it from being lazy on the way there.

The result and the bill

Six skins morphing one into the next, 2K widescreen, five Chinese voiceover lines plus the official theme, 26 seconds:

The five nominal clips add up to 25 seconds; the final container is about 26. I kept the full prompt, API payloads, and mixing command in a separate workflow and prompt manual instead of making this page one enormous terminal window.

The MiniMax account spent 35 RMB and still had 10 left. That total includes the finished clips, test generations, and an attempted music track. At the price at the time, the five 2K production clips themselves were roughly 20 RMB. The expensive line item is always “one more try should fix it.”

The Alibaba Cloud round spent more than 80 RMB without a publishable result; the 35 RMB MiniMax round produced the finished video. Those two bills cannot rank the providers forever. For this attempt, replacing the source art and spelling out the motion was cheaper than buying another draw: the first round bought a lesson, while the second at least left me with a finished video.