You've probably felt it: you generate a beautiful first shot of your main character, then the second shot gives you a slightly different nose, a different jacket, and hair that changed color somewhere between cuts. In an AI short video, that tiny drift breaks the illusion faster than almost anything else. Viewers may not name it, but they feel it — the person on screen stops being *a person* and becomes "a series of images."
The good news is that character consistency isn't luck. It's plumbing. If you set up your reference image workflow correctly, the same face walks through every shot, whether you're making a 30-second vertical clip or a multi-episode AI drama. Let's walk through how to do it inside 墨川导演台 (MiniMax H3 AI Director), step by step.
Step 1: Build a Character Reference Before You Generate Anything
Here's the analogy I use: a reference image is a passport photo for your character. Every border guard (every new shot) checks it. If the passport is blurry or the photo is from the wrong angle, you get stopped.
So before you touch a single storyboard card, create your character's "passport."
- Open the 资产库 (Asset Library) in 墨川导演台. This is where characters, scenes, props, and reference videos live.
- Create a character entry. Give it a name you'll actually remember — "Lin Wei — red jacket" beats "Character 1."
- Generate a clean reference using AI 制图 (AI Image). You can use text-to-image to create the base look, or image-to-image if you already have a photo or sketch you want to anchor to.
- Upload that image into the character's asset slot. This becomes the anchor the system uses for consistency across shots.
A few practical tips for a reference image that actually works:
- Front-facing, neutral expression, even lighting. Save the dramatic angles for the actual shots.
- One character per reference. Don't cram two people into one reference image — it muddies the signal.
- Include the outfit. If your character wears the same jacket in Episode 1 and Episode 4, bake it into the reference so you don't have to re-describe it every time.
- Generate a few candidates, then pick one. AI image generation is cheap; regret is expensive.
If you're running a series with a recurring cast, do this for every main character up front. Ten minutes of asset work saves you hours of re-rolling.
Step 2: Wire References Into Your Script and Storyboard
Now that your character has a passport, you need to attach it to every scene they appear in.
The workflow in 墨川导演台 runs through the 剧本库 (Script Library) and the 项目库 (Project Library):
- Use AI 编剧 (AI Screenwriter) to draft your six-part script and storyboard. It supports Chinese, English, Spanish, Arabic, and Japanese, and it writes the draft directly into your script library — so you're not copy-pasting between tools.
- Open the storyboard cards. Each card represents a shot.
- For every shot where your character appears, attach the character reference from your Asset Library. This is the moment where consistency is either won or lost.
- If a scene has multiple characters, attach each one's reference to the same card. Don't rely on text descriptions alone — the reference image is what carries the identity.
Think of the storyboard card as a film set. The reference image is the actor walking onto that set. If you forget to call them in, the model improvises, and improvisation is where faces drift.
A tip that pays off: batch your shots by character. If Lin Wei appears in shots 1, 4, 7, and 11, set up all her references in one pass. It's faster and less error-prone than jumping back and forth.
Step 3: Generate, Review, and Fix Drift Early
Now for the fun part — generating video. In the 项目库 (Project Library), you can run single-shot generation to test a look, or batch generation when you're confident in the setup. Cloud inference runs through ComfyUI workflows, which is what handles reference-image consistency, shot-to-shot continuity, and batch queuing.
Here's the review loop I recommend:
- Generate the first shot only. Don't batch yet.
- Check the face against the reference. Zoom in. Does the jawline match? The eye spacing? The hair color?
- If it drifts, don't tweak the prompt first — check the reference attachment. Nine times out of ten, the problem is a missing or wrong reference, not a bad prompt.
- Once shot one is solid, generate shot two. Watch for continuity between them — lighting, wardrobe, and pose.
- Only then go to batch generation for the remaining shots.
The temptation is always to fire off twenty shots and hope. Resist it. Two good shots tell you your pipeline is clean; twenty mediocre shots tell you nothing except that you have twenty problems.
A note on descriptions: your text prompt still matters, but it plays a supporting role. Use it to describe *action, framing, and mood* — "medium shot, walking through rain, tense" — and let the reference image carry *identity*. When you overload a prompt with physical descriptions, you're competing with your own reference.
Step 4: Handle Voice, Motion, and Multi-Language Projects
Consistency isn't only visual. If your character's face stays the same but their voice changes between scenes, the illusion still cracks.
- AI 配音 (AI Voice) uses IndexTTS 2.5 with multi-language support and emotion control. Lock in a voice per character in your 音频库 (Audio Library) the same way you lock in a reference image. Voice cloning and lip-sync live here too.
- For music-driven pieces, AI MV handles music video generation — useful for short-form content where the character needs to stay recognizable across fast cuts.
- For non-English projects, the AI Screenwriter's five-language support means your script and storyboard stay in one place, so your references don't get orphaned when you switch languages.
And if you're producing at volume, 云部署 (Cloud Deployment) matters. 墨川导演台 supports generic cloud GPU deployment across AutoDL, Alibaba Cloud, Tencent Cloud, 仙宫云, and custom SSH servers. It connects to a cloud ComfyUI over SSH, with local tunnel ports automatically isolated so multiple users don't collide.
If you want to test the cloud path on 仙宫云:
- Register at https://www.xiangongyun.com/register/B1H9M6 — you must register fresh and enter invite code B1H9M6.
- After registration and verification, send your 仙宫云 account ID to customer service (something like
ac401472-db02-5df3-xxxx-xxxxxxxxxxx). - Once your test account is opened, you'll see the Director's image — deploy an instance with it.
- When the instance is running, click SSH on the right side of the instance to get the host, port, and password.
- Enter those into the Cloud Deployment page in 墨川导演台 (SSH Host / SSH Port / SSH Password).
- Click Verify SSH, then Start Cloud ComfyUI, and wait for the green success indicator.
- Now write a script and test a video generation.
You can also run everything locally on your own GPU machine, or call the platform via API if you're integrating it into a larger pipeline. Specific hardware requirements and pricing vary — check with the official channels rather than guessing.
Wrapping Up: Consistency Is a Habit, Not a Trick
If you take one thing from this guide, make it this: character consistency in AI short video is 80% preparation and 20% generation. Build the reference, attach it to every card, review early, and fix drift before it compounds. Do that, and your audience stops noticing the seams — which is exactly the point.
A quick action plan:
- Open 墨川导演台 and build reference images for your main cast today.
- Run one test shot per character and compare against the reference.
- Once clean, batch the rest.
- For cloud setups, get your SSH details sorted before you're mid-project and frustrated.
- For anything unclear — account setup, deployment, feature questions — reach out directly.
You can find the platform at www.ymcdirector.com, add customer service on WeChat at ymcandai, and search 「杨墨川AI」 on Taobao. There's also a blog at blog.ymcdirector.com and support at info@ymcdirector.com.
Now go give your characters a face that survives the cut. 🎬
