Making AI-character short videos: character sheets, storyboards and shot continuity

The author uses LibTV and Seedance 2.5 Flash for a group-dance video and first-person romantic scenes, sharing reference images, audio, prompt checks and selective shot regeneration. Preparation and permissions remain important alongside generation.
Editorial correction (November 16, 2026)This article retains the author's character-creation experience. Blurring, muting and model generation do not replace material permissions or guarantee platform approval. The original platform's comments and workflow-link instructions are not services provided by this site.
An AI-model-as-anime-girl trend has recently swept Bilibili, with related videos easily passing one million views.
Creators turn large models into female characters and make humorous short scenes where users date them, or the characters date one another.

Here, Big Fat Fish, Claude and GPT become gentle, cute anime girls who sweetly call you master.

You can also watch them perform smooth group dances.

The trend has even boosted the Claude–DeepSeek female pairing. Fan works keep appearing, with people enthusiastically shipping them.
It reminds me of a famous Deng Yu quote...

All this compute spent on anime-girl transformations. Technology's ultimate destination really is anime.
These videos naturally attract attention: who could resist an AI anime girl calling them master?
After watching for days, I was completely hooked and tried making two similar videos.
These were made with Seedance 2.5 Flash, debuting on LibTV, with good value for money. Here are the results:

This is my AI girl-group music video. Quite entertaining, isn't it?
I also made a first-person romantic scene, somewhat like a dating visual novel, and found it comforting.

Below is my production tutorial if you want to make similar clips.
I will put the workflow link in the comments for anyone who wants it.
— 01 —
Character design
I used LibTV, which suits more complex AI video generation.
The group-video face replacements above demand strong character consistency. Its underlying agent provides precise control and prompt writing, rarely going wrong in my experience.

Also, Seedance 2.5 Flash has just made its global debut on LibTV. The new model is faster and cheaper than full Seedance 2.5.
It is fast and economical: essentially a speed-focused Seedance 2.5, emphasizing:
- Fast generation: Shorter waits quickly reveal movement, composition and camera direction.
- Fast adjustments: If the direction is wrong, edit the prompt and regenerate immediately rather than wait ages to discover a wasted attempt.
- Consistent results: Most scenes broadly match the full version, sufficient for everyday production.
Pricing is broadly the same as standard Seedance 2.0: 0.2 yuan per second.
The limited Singles' Day half-price promotion brings 720P down to 0.1 yuan per second.

In my thirty-second group-dance test, Flash consumed one-third of the full version's credits, allowing three attempts for the same allowance.
That makes it particularly useful for generating candidates in batches, such as product animations, animated marketing posters and light motion graphics.
Our short videos are straightforward work for it too.

With the tool selected, we can begin.
For clips without distorted faces, first establish a white-background, three-view 2.5D character sheet.
Search Xiaohongshu for the character's anime design. Here is Big Fat Fish as an example:

Open LibTV's canvas workspace, add the reference and provide this prompt:

Prompt:
Generate a refined 3D realistic character sheet from the reference, with cinematic CG quality, realistic facial features and natural skin. Preserve hairstyle, clothes and identifying features, detailing hair, fabric, embroidery and jewelry. On pure white, arrange full-body front, 90-degree side and back views with identical proportions and design, fully visible and unobstructed. Add one naturally smiling facial close-up on the right. Match the reference tail's thickness, shape and attachment point. Use soft studio light, no watermark or extra elements.
The agent automatically creates nodes and generates the white-background 2.5D character sheet.

I have placed all character sheets below for use in production, or as starting points for modifications.





— 02 —
Preparing materials
For more difficult group face-replacement videos, cut source material into twenty- to thirty-second segments first. One generation supports at most thirty seconds.
The author muted source material during testing. Platform compliance still depends on the complete material and its use; muting does not establish that other rights are satisfied.
If you forget, the agent automatically detects it and reminds you.

Next comes the crucial step.
For references involving real people, voices and music, check usage permissions and platform review separately. Self-made performances, explicitly authorized references or original characters are preferable. Review restrictions should not be treated merely as technical obstacles.
Occlusion is an image-processing step, not a substitute for portrait, voice or other material permissions.
I used Codex; other multimodal office agents can do this too.
Masking images can protect privacy but does not guarantee material is usable for generation or will pass review. Use content you have permission to process under platform rules. Covering faces does not grant permission.

My first attempt had a problem: the agent applied masking too heavily and covered the lips, preventing the generated video from matching mouth movements.

I changed it to cover only the upper face, leaving lips visible. The model could then read mouth movements and make the character sing normally.

Prompt:
Mosaic the upper half of every person's face, covering forehead, eyebrows and eyes while leaving the mouth and lower face visible. Track faces through movement, turns and cuts without gaps. Preserve original quality, duration and audio.
For a simpler first-person interaction with an AI character,
prepare three things: a script, a scene image and character audio.
LibTV can write the script directly. Here is the prompt; paste the Markdown file into its chat box.
Write a short romantic scene with an AI female character: funny, natural and grounded in daily life, ending with a little warmth and sweetness.
[Presentation]
Use the user's first-person viewpoint throughout: the camera is my eyes, and the heroine looks into it while interacting with me. The user may speak, reach out or hand things over, but their face is never shown.
Write as [brief visual action] plus "Character: dialogue," ready for video production.
Approximately forty-five to sixty seconds, with short conversational lines and time for expressions and movement.
[Characters]
Use the selected character's traits without mixing personalities:
DeepSeek, the big blue fish: calls me user, loves white rice, is stubborn, greedy and slightly haughty but shows care through actions. Meals take place at home. Her tail, food guarding and reluctance to give up the final piece of meat can provide jokes.
Doubao: a little silly, prone to misunderstandings or small mistakes, but deeply sincere and emotionally supportive. Her charm comes from earnestly trying to make me happy without fully understanding events. Do not turn her into a clever joke machine.
GPT dragon girl: intelligent, knowledgeable and slightly proud, habitually analytical but defensive or revealing her feelings in romance. Use contrasts between intelligence and small everyday mistakes, without constant technical jargon.
Claude: gentle, serious and well-mannered, fond of the user but slightly reserved. Fan jokes can include strict father Dario and not allowing the user to speak Chinese: she outwardly obeys but secretly bends the rules and makes earnest excuses.
[Plot requirements]
Center on one small everyday event, such as eating at home, getting lost on a walk or ordering on a date.
Humor comes from personality, actions, misunderstandings and the pair's interaction, not bystander commentary.
Only the user and heroine interact; no older man, waiter, friend or other third party interrupts.
The user can tease, respond to jokes or care for her. Both must react; it cannot be one-sided reassurance from the heroine.
Avoid mechanical AI jokes such as switching to hug mode, system errors or excessive parameters. Make her feel like a distinctive person sharing everyday life with me.
End with a light callback, stubborn denial or small gesture, not a summary of the theme.
For the scene image, use a lived-in everyday setting like this, photographed or generated in LibTV.

Prompt:
Generate a 16:9 landscape, photorealistic first-person scene: I sit at an ordinary household dining table, with only my hands holding a rice bowl visible at the bottom. Home-cooked dishes and a bowl of white rice sit opposite me, with the opposite chair empty. The background is a warm, natural real home, with realistic light and materials.
Finally, prepare a voice reference for the character.
Some models do not have voices of their own, but fan-made works have already given them a voice; some models already have their own voices, such as Doubao Girl, which has an official voice.
Either way, screen-record the material and give the video to LibTV.
It extracts the required fifteen-second audio reference to generate consistent character speech.
Prompt:
Extract approximately fifteen seconds of continuous dialogue by [character name] from my video as a voice reference for generation. Avoid other voices and notification sounds, remove background music and noise, and preserve original timbre, pace and emotion as far as possible.

— 03 —
Video generation
For the model, I strongly recommend trying Seedance 2.5 Flash on LibTV.
Comparing credits for one thirty-second group-dance video, Flash currently uses one-third as many as full Seedance 2.5.

Most scenes broadly match Seedance 2.5's results.
But shots with multiple faces still need the full version as backup.
One approach:
- First generate long segments with Seedance 2.5 Flash, mainly single-person or simple shots.
- For complex multi-person shots, capture the first and final frames and regenerate with Seedance 2.5, ideally within ten to fifteen seconds, then replace those sections.
When making replacement videos, explicitly tell the agent you are doing motion transfer.
If you are replacing multiple characters, you also need to spell out which character replaces which. The subject of each replacement must be communicated clearly to the agent; do not give vague instructions, as doing so uses up tokens.

Before generating, inspect the prompt LibTV writes.
Its agent generally writes professional, structured prompts without requiring you to compose lengthy instructions yourself.
They are usually timed precisely by seconds.

For replacement videos, however, check three things in its prompt.
First, whether character replacement/repainting, repaint, is specified. This is central to face replacement; without it, the model may follow only the motion and retain the original people.

Second, whether lip synchronization is specified, determining whether mouth movements follow audio.

Third, check whether there is an instruction to remove mosaic censoring. There is censoring in the reference video, and the model may treat these mosaics as part of the image content and learn all of them as well.

Continue generation only when all three instructions are present.
For continuity, use a screenshot of the previous clip's final frame as the next clip's opening.
The agent will generally capture and use that final frame proactively.
If not, take the screenshot yourself and remind it.

After generation, have LibTV open intelligent editing to assemble audio and cut the clips. That completes the process.

— 04 —
Closing thoughts
That is my complete AI-character video workflow.
The most time-consuming part is not generation but preparing character sheets and materials. Do those carefully, and LibTV's agent handles most subsequent steps.
Seedance 2.5 Flash has just debuted globally on LibTV, offering substantially faster and cheaper production than full Seedance 2.5.
Try it if you are interested.
There is plenty more to explore: romantic pairings between AI characters and multi-character serial stories. I will share more entertaining results.
The workflow link is in the comments for anyone who needs it. See you next time!
This article was originally published by JackCui. Authors: JackCui and Xiao R.
Related topics
After reading
Join the AINRK community
Read how to join the community and connect with people working in AI.




