Building an AI Short Drama Asset Library: Characters, Scenes, Props and Reference Videos
Here's a truth that took me too long to learn: the hardest part of making an AI short drama isn't the prompt. It's the memory.
You can write a beautiful six-shot script, hit generate, and get six beautiful shots that look like they came from six different universes. The detective has a new face in shot three. The café changes its wall color between cuts. The prop gun morphs into a phone. Every model is talented and every model has amnesia.
The fix isn't a better prompt. It's a better asset library. Think of it like a film crew's continuity binder — the folder where the script supervisor pins polaroids of every actor, every location, every prop, so that Day 1 and Day 12 still match. In 墨川导演台 (MiniMax H3 AI Director), that binder lives in the Asset Library, and building it well is the single highest-leverage thing you can do before you generate a single frame.
Let me walk you through how I set one up.
Why the Asset Library Is Your Continuity Department
Most AI video tools treat each generation as a fresh start. You paste a prompt, you get a clip, done. That works fine for a one-off landscape shot. It falls apart the moment you're telling a story across multiple shots — which is exactly what a short drama is.
The Asset Library in 墨川导演台 is built around a simple idea: assets are reusable, versioned, and shared across a project. Instead of re-describing your lead character in every prompt and hoping the model lands in the same place, you register them once as an asset, attach a reference image, and pull that asset into each shot. The model gets the same anchor every time.
This is where character consistency actually comes from. Not from magic words like "the same woman as before" — models don't remember. It comes from feeding the same reference material into every generation, and letting the workflow handle the rest.
The Asset Library organizes four kinds of assets:
- Characters — your cast, each with reference images
- Scenes — locations and their established look
- Props — objects that need to stay recognizable across shots
- Reference videos — motion, pacing, or style references you want to echo
If you're building a series, this is also how you keep Episode 1 and Episode 7 looking like the same show.
Step-by-Step: Building the Library
Step 1 — Start with characters, not scripts
Counterintuitive, I know. But your characters are the hardest thing to keep consistent, so build them first.
For each character, I create an asset and upload a reference image. A few practical notes:
- Front-facing, neutral lighting, plain background. A dramatic three-quarter shot with moody shadows looks cool but gives the model less to anchor on.
- One character per reference. Group photos confuse the model about who's who.
- If you need multiple angles, generate them first with AI 制图 (text-to-image or image-to-image), then upload the best ones as additional references for the same character asset.
The AI 制图 module is genuinely useful here — I'll prompt for a character sheet, pick the cleanest frame, and register it. That single image becomes the face your whole drama inherits.
Step 2 — Lock your scenes
Scenes are the second-biggest source of continuity drift. A living room that's warm and cluttered in shot one and sterile in shot four reads as a different apartment.
Same process: create a scene asset, attach a reference image of the established look. If your scene has a "hero angle" — the shot you'll return to most — use that as the reference.
A tip I've found helpful: name scenes by their story function, not just their look. "Lin's apartment — night, tense" is more useful than "living room 2." When you're assembling a shot later, the name tells you why the scene exists.
Step 3 — Register props that matter
Not every object needs to be an asset. A random coffee cup doesn't. But anything the audience will track — a locket, a specific phone, a signature jacket, a weapon — should be registered.
Props are the sneaky ones. They're small, so they feel unimportant, but they're exactly the kind of detail a model will happily reinvent. Register the ones that carry meaning.
Step 4 — Add reference videos for motion and mood
This is the step most people skip, and it's the one that separates "AI clips" from "AI drama."
Reference videos let you show the system the kind of movement, pacing, or visual rhythm you want. If you have a shot from a film whose camera movement you love, or a previous generation that nailed a gesture, you can keep it as a reference asset and pull it into future work.
Think of reference videos as choreography notes. You're not copying the shot — you're telling the model "move like this."
Step 5 — Organize by project
The Asset Library sits alongside the Project Library, which is where generation tasks and versions live. Keep your assets scoped to the project they belong to. A character from your sci-fi series shouldn't clutter your period drama.
If you're working across multiple dramas, this separation is what keeps your continuity clean.
Turning Assets into Shots
Once your library is populated, the actual generation gets much simpler — and this is where the workflow pays off.
In the Script Library, you manage scripts, episodes, and storyboard cards. Each storyboard card is a shot. When you generate a shot, you pull in the relevant character, scene, and prop assets, and the cloud workflow — running through ComfyUI with MiniMax H3 as the main model — uses those references to keep things consistent.
A few things worth knowing:
- Single-shot generation is for testing. Generate one shot, check the character's face and the scene's look, adjust, regenerate.
- Batch generation is for production. Once a shot's setup is proven, queue the rest of the episode.
- The AI Log / System Log shows you what's actually happening during generation. When something drifts, the log is where you find out why.
For consistency specifically, the reference images do the heavy lifting. The workflow supports reference-based consistency, so the more disciplined your library, the more stable your output.
If you're running on cloud GPUs — 仙宫云, AutoDL, Alibaba Cloud, Tencent Cloud, or your own SSH server — the Cloud Deployment page handles the connection. You fill in the SSH host, port, and password, verify the connection, then launch cloud ComfyUI. Local tunnel ports are isolated per user, so multiple people can work without stepping on each other.
One note on the 仙宫云 test flow: you need to register through the invite link and enter invite code B1H9M6, complete verification, then send your account ID to support to get access to the director console image. After that, the SSH details from your instance go straight into the Cloud Deployment page. Full details are on the official site — and if anything's unclear, the customer service WeChat ymcandai is the fastest way to sort it out.
A Few Habits That Save Hours
After building a few of these libraries, a few patterns emerge:
Build the library before the script, not after. It's tempting to write first and cast later. Don't. Your script should be written *for* the assets you have.
Be ruthless about reference quality. One clean, well-lit reference image beats five mediocre ones. The model averages what you give it.
Version your assets. When a character's look evolves, add a new reference rather than replacing the old one. You may need the earlier version for a flashback.
Use the AI Screenwriter for drafts, then refine. The conversational screenwriter generates six-part scripts and storyboard drafts in Chinese, English, Spanish, Arabic, or Japanese, and writes them into your script library. It's a great starting skeleton — then you cast it against your assets.
Keep the audio side in mind. The Audio Library handles voice cloning, lip sync, background music, and sound effects, and AI Dubbing (IndexTTS 2.5) supports multilingual voiceover with emotion control. If your character has a fixed voice, treat it as part of the asset too.
Where to Go From Here
An asset library is unglamorous. Nobody posts a screenshot of their character reference folder. But it's the difference between a demo and a drama — between six disconnected clips and a story that holds together.
Start small. Build one character, one scene, one prop. Generate a two-shot sequence and see how much more stable it feels. Then scale up.
If you want to see the full workflow — script, storyboard, assets, generation, dubbing — the official site at www.ymcdirector.com has the details, and the blog at blog.ymcdirector.com goes deeper on specific features. For hands-on questions, the customer service WeChat is ymcandai, and you can also find the team by searching 「杨墨川AI」 on Taobao or 「杨墨川」 on Douyin, Xiaohongshu, and WeChat Channels.
Build the binder first. The shots will thank you.
