AI Idol Dramas Fear Face Swapping; Create Character Cards Before Demos

Last week, I read a Huxiu article saying domestic AI is filling the global idol drama void, with transfer students speaking Sichuan accents and Northeast girls confronting mafia heiresses. My first thought was: Can the character keep the same face from start to finish?

AI short dramas mean letting AI turn scripts into animated videos. I've attended many hackathons and seen countless demos. Version one is explosive, version two has a sharper chin for the protagonist, and version three sees school uniforms turn into suits. Viewers drop out for simple reasons: they don't recognize the person from earlier. My approach is to create a Character Bible first. In plain terms, it's fixing the character into a replicable setting card, including face, hairstyle, clothes, verbal tics, and what absolutely cannot change.

Run Two Lines First to See Differences

I tried a cliché template: A transfer student enters a Korean chaebol campus and meets three school bullies. Beginners can copy this directly.

Line one: I input a sentence into a video generation tool: "Korean campus, cute transfer student with Sichuan accent, three bullies watching, anime style, vertical screen." This is text-to-video, giving only text and letting AI draw. On my end, I saw a short preview in about twenty minutes. But close-ups changed the face; wide shots changed the clothes.

Line two: These past few days, I've been trying Notion Database as a Character Bible. Notion is a note/table tool, and Database is a filterable table. Open Notion, click New Page, title it Character Bible. In the body, type /database and select Table View. Create columns: Name, Face Shape, Hairstyle, Outfit, Verbal Tic, Forbidden Changes, Reference Images. For the transfer student row, write round face, black hair with bangs, blue-white uniform, slight Sichuan tail-end tone, forbidden to change face shape, forbidden to change uniform.

Then I used text-to-image tools to generate three reference images: front half-body, side full-body, and back view. Text-to-image means words becoming pictures. Tools like Nano Banana 2 work here. The key is that all three images must be the same person. After getting the images, go into domestic video tools like Kling AI, Jimeng AI, or Vidu, find Image-to-Video, upload the first frame, write a short prompt: maintain round face and uniform, slow head turn, background light spots, vertical screen. Image-to-video is more stable than text-to-video because AI can follow the image.

On my end, Line two took longer, about fifty minutes. But I could edit out a watchable demo. Line one required reworking storyboards and character sheets, ending up taking similar time. The difference is Line two gives peace of mind.

Where Beginners Most Often Crash

The biggest pitfall is character faces breaking across shots. Materials repeatedly mention this: AI comic dramas start with round faces and end with oval faces, causing viewers to leave immediately. The solution is deletion. Keep only necessary features in prompts, clearly state in the Character Bible: don't change face shape, don't change hairstyle, don't change main outfit color.

Second pitfall: Accents. AI voiceover allows choosing timbre, but accent isn't tone. Relying solely on voiceover for a Sichuan accent might turn into Mandarin. My method is to fix two short phrases in the verbal tic field, like "Mei de wenti ma" (No problem) or "Yao de ma" (Okay), ensuring they repeat in dialogue.

Third pitfall: Trying to make a whole series at once. Beginners often start from Episode 1, resulting in a finished piece full of variables. I usually start with 15 seconds, one character, two actions, one shot. Once stable, expand to classroom, hallway, and rooftop scenes.

Another pitfall is platform queuing. AI video tools often require waiting, especially during evening peaks. Don't start ten tasks at once; put scripts, character images, and prompts into Notion first. When the queue gets long, go edit the next segment.

After learning this, try switching three scenes for the same character. Change only background and action, not character description. If this works, multi-character frames and vertical storyboards will be much more stable.

AI idol dramas ultimately compete on who can lock down the character.