Dayward AIWeChat Mini Program
Build an AI Short-Drama Production Pipeline With Agents in 14 Days
Week 1 · D3🔒 LockedAbout 5 hours

Character Consistency: Character Sheets, Reference Images, and Style Locking — Keeping the Same Person the Same Person in Every Shot

Wire up an image generation API, batch-produce character sheets and prop/scene assets from the shot data, lock down a look with reference images and fixed templates, and store the assets in a reusable library.

Today's goals 0/3

Sign in to tick these off and save your progress.

Today's Goals

  1. Wire up an image generation API, and batch-produce character sheets and scene images from the shot data
  2. Explain what character consistency relies on at the API level, and lock down a character's look with a reference image
  3. Build the generated assets into a searchable asset library, so later shots reuse them directly instead of regenerating

Yesterday you obtained structured shot data: who is in which shot, what they say, what is in frame. Today turns those fields into something genuinely visible — images. Come back and tick off the three goals.

Plain-Language Walkthrough

The lead has a different face in the second image

Start with the scene everybody hits the first time they make an AI short drama.

You write the prompt: a man of twenty-five, short black hair, heavy brows, a blue delivery uniform. Generate the first image and you are pleased. Generate a second with the same description and only the action changed to standing at a convenience store door — and out comes a different chin, wider-set eyes, and a blue that is not the same blue. You make the description finer, adding a faint scar at the left brow, and the third face changes again with the scar moved to the right.

The endpoint is not broken and your prompt is not insufficiently good. The reason is plain: an image generation model has no memory. It does not know that a previous image exists, and each request is a brand-new, unrelated sample.

The crew analogy makes it clear. You hired an extremely fast concept artist producing a look test a minute, with one quirk: they forget the previous one completely on finishing it. Say "the same as that last one from another angle" and they ask which one. The only thing they can work from is the sheet you hand them at this moment — they draw what is written and invent whatever is not.

That invented part is where inconsistency comes from. Your description covered the hairstyle and the uniform and not the nose bridge, the face's width, the eyelid, or the uniform's exact shade — those unconstrained degrees of freedom get re-rolled with every sample. And a face's recognizability lives precisely in those details: an audience does not remember a blue uniform and recognizes in two seconds that this is not the same person.

The engineering cost is larger than it looks. An episode of forty shots has the lead in thirty of them. Rerolling to catch a matching face costs, at ¥0.025 an image and ten rolls per shot, only ¥7.5 — money is not the problem, people are: you must sit comparing image after image and rerolling, and thirty shots is an afternoon. This pipeline's goal is running an episode unattended, and picking images by hand is its greatest enemy. Worse, that cost grows linearly with episodes: five episodes is five afternoons and there can be no second season.

So character consistency is not an aesthetic problem but the precondition for this line being automatable at all. So what, at the API level, holds one person in place?

Sign in and link WeChat to read

Sign in, then link your WeChat mini program account on the account page. Linked accounts share one free article per day. Unlock additional articles with rewarded videos in the mini program.

Mini program QR code

Scan in WeChat to open Dayward AI. Open the same course and chapter, choose the rewarded video, then refresh this page after completion.

Interview questions

  • Where does the character consistency problem in image generation come from, and what engineering mitigations exist, with what trade-offs?生成模型的角色一致性问题是怎么来的?工程上有哪几种缓解手段,代价分别是什么?
    Common in ChinaCommon overseasBasic#image-generation#consistency

    🔒 Unlock this chapter to see the answer keys.

  • Does fixing the random seed solve character consistency? What does a seed actually lock?固定随机种子能解决角色一致性吗?它到底锁住了什么?
    Common in ChinaCommon overseasIntermediate#image-generation#reproducibility

    🔒 Unlock this chapter to see the answer keys.

  • How would you design the cache key for reusing generated assets so that you save money without serving the wrong asset?生成类资产要做复用,缓存键你会怎么设计,才能既省钱又不会串戏?
    Common in ChinaCommon overseasDeep dive#caching#cost#image-generation

    🔒 Unlock this chapter to see the answer keys.

Comments