**English** | [δΈζ](../zh/character-reference-lock.md)
[β All prompt templates](../../../README.md#-prompt-templates-by-category) Β· [Template index](./README.md)
# π§± Reference image identity lock
> Name every reference with a stable token, enumerate what to inherit from it, and separately enumerate what must not be inherited. The inherit-nothing-else clause is what separates working locks from broken ones.
## Copy this
Hit the copy button on the block, replace everything in [square brackets] with your own details, and send it to any AI chat (ChatGPT, Claude, Gemini) together with your reference images. It will write a Seedance-ready prompt for you that follows this template.
````text
I want a clip where the character looks the same from start to finish. [I am sending you 2 reference images: a young woman with short hair and thin-framed glasses.] [What she does: browses a bookshop, looks up and smiles at the camera.] Using the prompt template below, rewrite it into one ready-to-use Seedance video prompt for me:
#### Reference image identity lock
Name every reference with a stable token, enumerate what to inherit from it, and separately enumerate what must not be inherited. The inherit-nothing-else clause is what separates working locks from broken ones.
**Use when:** Any clip where a face, an outfit, a product or a UI layout must survive across shots. Applies to Seedance 2.0 and 2.5 alike; 2.5 additionally accepts audio and video references under the same token scheme.
**Guidance:**
- Split references by role and lock each separately. The GoPro fishing case declares `@location1` for the river and `@hands1` for the forearms, tools and bottle, each followed by `100% matches reference`.
- Enumerate the inherit list instead of writing keep her consistent. The boyfriend-POV case lists thirteen items: identity, features, face shape, skin tone, apparent age, hairstyle, hair colour, height, build, body proportion, clothing, footwear, overall bearing.
- Always add the do-not-inherit clause. Without it the reference's background, pose and lighting come along; the anime duel case spells out that the reference's background, room, furniture, text, split layout, pose, angle and framing must not be reproduced.
- For a UI or scene reference, separate composition lock from design lock. The character-select case marks `@image1` as LOCKED SCENE COMPOSITION and `@image2` through `@image6` as design-only, then adds DO NOT reproduce their reference poses.
- Identity drifts hardest during head turns, occlusion and fast motion. List those poses explicitly and restate the same-face requirement for the high-energy segments.
**Examples:** [#1](https://goodcase.ai/cases/elsasofia-ai-seedance-ai-e087ab2aed4b) [#2](https://goodcase.ai/cases/liyue-ai-seedance-ai-dd263958ed42) [#3](https://goodcase.ai/cases/case-79acf1a3e8a6) [#4](https://goodcase.ai/cases/seedance-2-5-ui-228cf63ce8ff)
**Structure:**
1. Token declaration: give each reference a name β `@image1`, `@Image2`, `<<>>`, `@location1`, `@hands1` β and reuse it verbatim everywhere
2. Inherit list: enumerated attributes pulled from the reference (face shape, features, hair colour, body proportions, wardrobe items, accessories)
3. Do-not-inherit list: background, room, furniture, pose, composition, framing, original lighting, any text
4. Cross-shot clause: same face when turning, looking down, speaking, or with a hand near the face
5. Negative: no cloning, no duplicates, no feature averaging, no attribute swaps between characters
**Pitfalls:**
- Uploading a reference without any textual lock. The model then treats it as a style reference, not an identity reference.
- Compressing wardrobe into same outfit. Every working case in the corpus breaks the outfit into individually named garments and accessories.
- Feeding a sketch or illustration reference without a render instruction. Add use only as design blueprints, render as fully realistic live-action humans, or the line-art survives into the video.
- Adding references that no token names. Each unnamed extra image is one more chance for attributes to bleed between subjects.
````
## Three steps
| Step | What to do |
| --- | --- |
| 1 | Copy the whole block above and replace the [bracketed] parts with your product, person or scene. Attach reference images if you can. |
| 2 | Send it to any AI chat and get back a Seedance prompt written to this structure. |
| 3 | Paste that prompt into Seedance (Dreamina / Jimeng) and generate. If the result is off, check the pitfalls first, then adjust and re-run. |
## Cases in this category (5 filed, by heat)
| Preview | Case | Version | Heat |
| --- | --- | --- | --- |
|
| [Seedance 2.5 Multi-Reference Character Lock Short](https://goodcase.ai/cases/elsasofia-ai-seedance-ai-e087ab2aed4b) | 2.5 | 64 |
|
| [Seedance 2.5 Boyfriend-POV Couple Short with an Authentic Handheld Phone Look](https://goodcase.ai/cases/liyue-ai-seedance-ai-dd263958ed42) | 2.5 | 30 |
|
| [Seedance 2.5 Spider Anti-Hero Character Select UI Animation](https://goodcase.ai/cases/seedance-2-5-ui-228cf63ce8ff) | 2.5 | 8 |
|
| [High-Quality Anime Swimsuit Video](https://goodcase.ai/cases/case-79acf1a3e8a6) | 2.0 | 5 |
|
| [4K Seedance 2.0 - Reference to Video](https://goodcase.ai/cases/4k-seedance-2-0-reference-to-video) | 2.0 | - |
---
[β Previous: Second-by-second timeline script](./timeline-shot-script.md) Β· [Next: Storyboard grid to video β](./storyboard-grid-to-video.md)