An identity lock prompt is a fixed block of text that describes who or what must stay the same (a face, an outfit, a product, a camera set-up), pasted word for word into every prompt, next to a separate shot block that changes each time.
This page is the template library: fill-in blocks for faces, wardrobe, objects, camera, group scenes and image-to-video, plus what each tool calls its reference-image feature. For why characters drift in the first place, read character consistency in AI.
Think of a casting card stapled to every page of a script. The card never changes: same name, same face, same coat. The page under it changes: new room, new action, new angle. An AI model forgets your character between images, so you hand it the same card every time. If you reword the card, even slightly, the model reads it as a slightly different person.
- Split every prompt in two: an identity block you never edit, and a shot block you rewrite each time.
- Same words, every time. No synonyms, no reordering.
- Use a reference image when the tool takes one. Text alone is the weaker lock.
- With a reference, name its job (“the person in image 1”) instead of re-describing it.
- Check each result against your first good image, not against the last one.
How the two-block method works
The plain point: one block holds what must not change; the other holds what this shot is about.
Vendors who publish prompting guides give the same advice in their own words. Black Forest Labs, which makes FLUX.2, says to define a character in detail and keep that description across panels, and to repeat the details in every panel prompt.1 OpenAI’s guide for its image models says to request one change at a time and repeat the details to preserve.2
Rules for writing the identity block:
- Concrete over flattering. “Short black bob, fringe to the eyebrows” locks; “beautiful hair” does not.
- Five to ten traits. Enough to be recognisable, few enough that the model can keep them all.
- Never reword it. Keep it in a notes file and paste it. “Navy coat” in one shot and “dark blue jacket” in the next are two different coats to a model.
- Keep it out of the shot block. The shot block is setting, action, light and camera only.
The identity block is pasted, never typed. Everything that changes lives in the shot block.
Face and character lock
The plain point: describe the face as a list of fixed traits, then end with a no-change line.
If you have a reference photo, add one line at the top: “The character is the person in image 1.” Keep the identity block anyway; it tells the model which traits matter most.
Outfit and wardrobe lock
The plain point: clothes drift as much as faces. List each garment with colour, material and one detail.
Changing outfits mid-story? Write LOOK B as a second block and say in the shot block which look applies.
Object and product lock
The plain point: with a reference image, name the image’s job and list what must not change. Without one, spell out shape, colour and label.
OpenAI’s guide says to identify each input image by number and purpose, and its product example asks to preserve the product’s geometry and label exactly.2 No vendor guide we read says outright “do not re-describe a product shown in a reference”, but two point that way. Google’s Gemini Omni docs say overly descriptive editing prompts can cause unintended changes.3 Kling’s guide says a subject with a voice already bound to it does not need its voice tone set again in the prompt.4 So the template below points at the reference and lists only what must survive.
With no reference, the text must carry everything. FLUX.2’s guide suggests exact hex colours for each part when brand colours matter.1
Camera lock
The plain point: for a series that must match, or a video shot that must not move, fix the camera in its own block.
Google’s Veo 3.1 guide lists camera movement, composition and lens as separate parts of a prompt, which is why this block keeps them apart.5 More camera words: camera grammar.
Scenes with two or more characters
The plain point: one identity block per person, a label for each, and positions in the frame.
Tools limit how many people they can hold. Google’s developer docs say its Nano Banana models mix up to 14 reference images, with a smaller limit for characters that varies by model.6 Kling’s 3.0 guide lists “Multi-Character Coreference (3+)” among its features.4
Image-to-video identity
The plain point: the image already holds the identity. The prompt should describe motion and camera, and say that nothing else changes.
Google’s Veo docs say the input image becomes the first frame, and that Veo 3.1 takes up to three reference images of a person, character or product to keep its appearance.7 Gemini Omni’s docs suggest adding “Keep everything else the same” when editing one aspect of a video.3
Turning a photo of a real person into video? Read the consent section of how to turn a photo into a video first.
In image-to-video, do not describe the face again. Describe the movement, and lock everything else.
A JSON-style template (for tools that read it)
The plain point: some models accept a structured prompt with labelled fields. Black Forest Labs documents this for FLUX.2.
Its guide says FLUX.2 reads structured JSON prompts and understands them as well as plain text, with fields such as scene, subjects, style, lighting, composition and camera.1 The identity block becomes the “subjects” entry, copied unchanged; the rest is the shot.
OpenAI’s guide says plain paragraphs and JSON-like structures can express the same intent, so this is a choice of format, not a stronger lock.2
What each tool calls its reference feature
The plain point: the name changes from tool to tool; the job is the same.
| Tool | Called | What the vendor says |
|---|---|---|
| ChatGPT | Called: a reference image attached to the message. | Vendor: attach it when you want ChatGPT to edit it or use it as visual guidance.8 |
| Gemini / Nano Banana | Called: reference images; the app lists “character consistency”. | Vendor: keeps a character’s appearance across images; separate limits for objects and characters (see above).9 |
| Veo 3.1 | Called: “ingredients to video”; reference images. | Vendor: reference images of a scene, character, object or style keep a consistent look across shots.5 |
| Gemini Omni | Called: subject reference. | Vendor: generates video with specific subjects given as reference images.3 |
| Kling 3.0 | Called: Elements (“Bind Subject”). | Vendor: upload 2-4 reference images of the element to keep the main character consistent.4 |
| FLUX.2 | Called: multi-reference editing. | Vendor: describe each input’s role, e.g. subject from image 1, style from image 2.1 |
Midjourney’s and Runway’s help pages blocked our reader on 12 Oct 2026, so their reference features are not described here. How to use reference images covers the technique in general.
Questions people ask
What is an identity lock prompt?
A fixed description of a character, outfit or object that you paste unchanged into every prompt, so the model keeps drawing the same thing. It works best alongside a reference image.
Do I still need a reference image?
If your tool takes one, yes. Google, for example, describes Veo’s reference images as the way to keep a person’s or product’s appearance in a video.7 The text block then tells the model which traits matter.
Why does the face still change?
Each image is drawn fresh, and OpenAI notes its models can still struggle with recurring characters.10 Compare each result with your first good image, and regenerate from that image rather than from a drifted one.
Can I lock a real person’s face?
Only with their permission, and some tools refuse. Google’s Gemini Omni API says uploading and editing images of certain recognisable people is not supported.3 See voice and likeness law.
Where to go next
- “Why do characters drift at all?” Read character consistency in AI.
- “I need to change one thing and keep the rest.” See image edit prompts that change only one thing.
- “What are seeds and LoRAs?” See seed, LoRA and image-to-video.
- “How do I plan the shots?” Try plan before you generate.
- Related: How to Keep the Same Character in a Comic or Picture Book With AI.
- 1Black Forest Labs, FLUX.2 prompting guide, undated, read at source 12 Oct 2026“The key is to define your character in detail and maintain that description across panels.” and “Repeat these details in every panel prompt.” ↑a“assign exact hex colors to each part for brand consistency.” ↑b“FLUX.2 interprets structured JSON prompts, giving you precise control over every aspect of your image.” and “FLUX.2 understands both formats equally well”. ↑c“subject from image 1, style from image 2, background from image 3.” ↑d
- 2OpenAI, Image prompting, undated, read at source 12 Oct 2026“Pass the previous output as the next edit input, request one change, and repeat the details to preserve.” ↑a“Identify each input by number and purpose: subject, style, clothing, or background.” and “Preserve product geometry and label legibility exactly.” ↑b“Short prompts, descriptive paragraphs, JSON-like structures, instructions, and tags can all express the same intent.” ↑c
- 3Google AI for Developers, Generate and edit videos with Gemini Omni Flash, undated, read at source 12 Oct 2026“Simple prompts work best for video editing. Overly descriptive prompts can lead to unintended changes.” ↑a“When editing a specific aspect of the video, include ‘Keep everything else the same’ to maintain visual consistency.” ↑b“You can generate a video incorporating specific subjects provided as reference images.” ↑c“Uploading and editing images containing certain recognizable people is not supported.” ↑d
- 4Kling AI, Kling VIDEO 3.0 Model User Guide, 6 Feb 2026, read at source 12 Oct 2026“it’s not recommended to set the tone again in the prompt.” ↑a“Multi-Character Coreference (3+)”. ↑b“Upload 2-4 reference images of the element.” ↑c
- 5Google Cloud, The ultimate prompting guide for Veo 3.1, 16 Oct 2025, read at source 12 Oct 2026“Camera movement: Dolly shot, tracking shot, crane shot”, “Composition: Wide shot, close-up”, “Lens & focus: Shallow depth of field”. ↑a“Provide reference images of a scene, character, object, or style to maintain a consistent aesthetic across multiple shots.” ↑b
- 6Google AI for Developers, Nano Banana image generation, undated, read at source 12 Oct 2026“Nano Banana models allow you to mix up to 14 reference images.” Per model, for example: “Up to 4 images of characters to maintain character consistency”. ↑
- 7Google AI for Developers, Generate videos with Veo 3.1 in Gemini API, undated, read at source 12 Oct 2026“Veo uses the input image as the initial frame.”, “Veo 3.1 now accepts up to 3 reference images to guide your generated video’s content.” and “Provide images of a person, character, or product to preserve the subject’s appearance in the output video.” ↑a“Provide images of a person, character, or product to preserve the subject’s appearance in the output video.” ↑b
- 8OpenAI, ChatGPT Learn, Image generation, undated, read at source 12 Oct 2026“Attach a reference image to the composer when you want ChatGPT to edit it or use it as visual guidance.” ↑
- 9Google, Nano Banana 2 – Gemini AI image generator & photo editor, undated, read at source 12 Oct 2026“Maintain a character’s appearance across generated images.” ↑
- 10OpenAI, Image generation, undated, read at source 12 Oct 2026“the model may occasionally struggle to maintain visual consistency for recurring characters or brand elements”. ↑
The comparison on this page (the casting card), the two-block method as written here and all templates are this site’s own, not claims from these sources. None of the templates carries a recorded test date.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK