To keep an AI character the same, write an identity block once, paste it unchanged into every prompt, change only the shot block, and give the tool a reference image wherever it takes one. Copy-paste identity lock templates for face, outfit, product and camera are just below. The rest of the page covers seeds, reference images (IP-Adapter, Higgsfield, ChatGPT Images 2.5), each tool’s own reference feature, multi-character scenes, turnaround sheets and identity lock lines for video.
Identity Lock Templates: Copy, Fill In, Reuse
The plain point: drift often starts when the description changes from one prompt to the next. Split every prompt in two: an identity block you write once and paste word for word every time, and a shot block that changes. Do not “improve” the identity block halfway through a series: new words describe a new person. The full library, with more variants, is on identity lock prompt templates.
Think of the identity block as the casting sheet pinned to the studio wall. Every scene changes; the actor does not. Fill in the brackets once, then never edit them for that character.
How to put them together: identity block, then outfit block, then product block if there is one, then camera block, then the shot block, in the same order every time. For video, the camera block can describe a move instead of a locked-off shot; keep it word for word when two clips must cut together.
Reuse the identity block word for word. Only the shot block changes between images. Add a reference image on top whenever the tool accepts one.
Why Character Consistency Is Hard
AI models generate each image independently. They don't remember the character they produced in a previous run — every generation starts fresh. Without explicit anchoring techniques, the same prompt produces a subtly different-looking person every time. Across a content series, that drift becomes obvious and the illusion of a consistent persona breaks down.
The good news: most major image and video tools now let you supply a reference image (see reference features by tool below). You have several methods available depending on your workflow and the tools you use.
Models forget every character. Each image is generated fresh, so without an anchor the same prompt gives a slightly different person each time. The usual anchor is a reference image.
Method 1: Seed Locking (All Image Models)
Every image model uses a random seed to start generation. The same prompt with the same seed, on the same model and settings, usually produces the same output. Find a seed that produces a subject with exactly the features you want, lock it, and vary only the environment, lighting, and outfit. The face tends to stay close across the series.
How to find the right seed: generate 10–15 variations of your character prompt. Note the seed from the best output (shown in generation metadata on most platforms). Lock that seed for all subsequent content in the series. To change the environment, update the environmental elements in the prompt while keeping the seed constant.
Limitation: Seed locking only works reliably within the same model and same base prompt. Moving to a different model or significantly changing the prompt structure breaks seed-based consistency.
Lock the seed, change the scene. Find a seed that gives the right face and vary only environment, lighting and outfit. It breaks if you switch models or rework the prompt.
Method 2: Reference Images (IP-Adapter + ControlNet)
IP-Adapter is a reference-image technique for open-weight model pipelines. Upload a reference image of your character and the model reproduces the identity at a weight you control. IP-Adapter Face ID variant specifically locks facial identity and is more precise than standard IP-Adapter for portrait work.
Black Forest Labs’ hosted FLUX.2 [pro] is not an IP-Adapter pipeline: its documentation does not mention IP-Adapter, and its editing endpoint takes reference images directly.1
Set IP-Adapter weight between 0.6–0.75 for the best balance of identity preservation and creative flexibility. Below 0.5 the model partially ignores the reference. Above 0.85 it copies too literally and output looks stiff.
These IP-Adapter numbers are working starting points, not vendor figures, and the right weight varies by model and adapter; how to use reference images suggests a slightly different range. Test two or three values on your own reference.
Use IP-Adapter Face ID for portraits. A weight of 0.6–0.75 balances identity and flexibility; below 0.5 the reference is partly ignored, and above 0.85 output looks stiff.
Method 3: ChatGPT Images 2.5 Conversational Consistency
ChatGPT Images 2.5 is the current image model in ChatGPT.4 It maintains context across a conversation, which suits creators who don't want to manage seeds or technical pipelines. Upload a reference image of your character, then describe changes while explicitly instructing it to maintain identity.5 For more prompts that change one thing and keep the rest, see image edit prompts that change only one thing.
The same conversational approach works with Nano Banana 2.1, Google’s update to Nano Banana 2, where editing is also a conversation.6 Generate your base character first, then use that image as a reference input for variations. It does not need the technical setup of a ControlNet pipeline. An earlier version of this paragraph also named Happy Horse here as an image model; it is a video model.
Method 4: Higgsfield Studio — Reference Photo Identity Lock for Video
Higgsfield Studio takes a face reference photo and uses it to hold identity across the generated clip. Expect the match to be strongest in simple shots; fast motion, extreme angles and busy scenes still test it, so check every clip. An earlier version of this paragraph called it the strongest identity consistency tool for AI video in 2026 and said it keeps exact facial identity regardless of motion, camera angle or environment; no source supports either.
For video identity, give the model a face. A reference photo in a tool such as Higgsfield Studio holds a face better than text alone, but check each clip, especially fast or busy ones.
Reference Features by Tool
The plain point: most tools now take a picture of your character as well as words, but each calls it something different and allows a different number. This table lists only what each vendor documents, read on 12 Oct 2026; features change, so check the linked page before you rely on a number.
| Tool | What the feature is called | How many references |
|---|---|---|
| ChatGPT Images 2.5 (OpenAI) | Feature: upload an image and describe the change; a selection tool edits one area.5 | Number: not stated on OpenAI’s help page. |
| Nano Banana (Google Gemini) | Feature: reference images, mixed in one prompt. | Number: up to 14 in total. Of those, up to 4 character images on Nano Banana 2.1 and Nano Banana 2, up to 5 on Nano Banana Pro, and none on Nano Banana 2 Lite.7 |
| Midjourney V8.1 and V8.2 | Feature: the Edit model (“Attach to prompt”), which replaces Omni Reference and Character Reference. On V7, Omni Reference (--oref, weight --ow) still applies.8 | Number: up to 4 with the Edit model; one image with V7 Omni Reference.9 |
| Kling 3.0, 3.0 Omni and O1 | Feature: Element Library: a saved character, prop, costume or scene built from several angles, which can also carry a voice. Kling says Kling 4.0, which takes more references, rolled out to all users on 15 October 2026.10 | Number: up to 4 images per element (1 main and 3 supplementary). In Video 3.0 Omni, up to 7 images or elements per video, or 4 when the input includes a video.11 |
| Veo 3.1 (Google, Gemini API) | Feature: reference images of a person, character or product, to keep its appearance in the clip. Clips that use them must be 8 seconds long. | Number: up to 3, on Veo 3.1 models only.12 |
| Seedance 2.5 (ByteDance, via Dreamina) | Feature: multimodal inputs (images, video, audio), referred to in the prompt by tag, such as @Image 1. | Number: up to 50 inputs in total.13 |
| Runway Gen-4 References | Feature: References for images: tag a reference to save it and call it by name with @; then animate the result with Runway’s video model. | Number: up to three per generation.14 |
Higgsfield’s reference photo workflow is covered in Method 4 above; this page has not read a Higgsfield document on the number of references it takes.
Use the tool’s own reference feature, then the identity block on top. The limits differ: three references in Veo 3.1 and Runway, four in Midjourney’s Edit model, up to 14 in Nano Banana.
Method 5: Pre-Production Storyboards (Lock Before You Generate)
The most underused consistency technique is the one that comes before generation: the pre-production reference board. Generate a Character Design Sheet first — a 6-panel board showing your character from multiple angles, expressions, lighting conditions, and with color swatches. Use this board as the reference input for all subsequent image and video generation.
This approach works across every model because you're providing a comprehensive reference rather than relying on any single seed or technical pipeline. It's the method used in professional animation and VFX production — adapted for AI workflows.
Turnaround sheet prompt: front, three-quarter, side and back
When you only need the angles, a four-view turnaround is quicker than the full sheet, and it gives a video tool a reference for every side of the character. Some tools can also make extra views for you: Midjourney documents asking its Edit model for new views of a character,9 and Kling says its O1 Image model can generate additional views from a single main reference image.11
Build a character design sheet first. A 6-panel board of angles, expressions, lighting and colour swatches becomes the reference for every later image and video, on any model.
Video Consistency: Identity Lock Text Instructions
For text-to-video models without reference image support, identity lock instructions are your primary tool. Add these to the [SUBJECT] section of every video prompt:
No reference support? Use identity lock text. Add a line to the [SUBJECT] section of every video prompt asking for the same face, hair and clothing, with no morphing.
Multi-Character Scenes: Keeping Two or More People Consistent
The plain point: give every character a name, their own identity block and, where the tool allows, their own reference image; then say exactly who stands where. Features bleed between people when the prompt lets them share words.
- Name each character and use that name every time, never “the woman” in one line and “she” in the next.
- Keep the identity blocks separate. One block per person, never one list of mixed features.
- Place them: on the left, on the right, in the foreground, facing each other.
- Match references to names, in order. Seedance’s own example tags each input, “Use @Image 1 for the venue. Reference @Image 2 for the pianist.”13 Kling says that in multi-character scenes its model can lock each character’s features separately.11
- If features still mix, Midjourney suggests combining the characters into a single reference image.9
One name, one block, one reference per character. Then fix their places in the frame and tell the model not to swap features between them.
AI Influencer Pipeline: Full Consistency Workflow
For AI influencer content production, the recommended full consistency pipeline in 2026:
- Step 1 — Persona generation: APOB AI ↗ for hands-on customized AI personas with full creative control, or Glam AI ↗ for template-driven beauty content where you provide a reference and it handles the prompting automatically
- Step 2 — Reference library: Generate 15–20 reference shots using the Character Design Sheet prompt. Different lighting, angles, expressions.
- Step 3 — Image content: ChatGPT Images 2.5 (conversational iteration), Nano Banana 2 Lite (volume/speed)15 — both using reference images from Step 2
- Step 4 — Video content: Higgsfield Studio for character-locked cinematic video using reference from Step 2
- Step 5 — Voiceover: ElevenLabs AI Studio ↗ — voice cloned once from a sample, used across all content
Questions people ask
How do I keep the same character in AI images and video?
Paste the same identity block into every prompt and add a reference image of the character wherever the tool accepts one. A single good reference can be enough: Runway says its Gen-4 References produce consistent characters across lighting and locations from just one reference image.14
What is an identity lock prompt?
It is a fixed description of a character (face, hair, outfit, distinguishing marks) plus an instruction to keep them unchanged, pasted word for word into every prompt in a series. The term is this site’s working name for the method, not a vendor feature; the templates are on identity lock prompt templates.
How many reference images can I use?
It depends on the tool: up to 3 in Veo 3.112 and in Runway Gen-4 References, up to 4 in Midjourney’s Edit model,9 and up to 14 in Nano Banana, of which 4 or 5 can be characters.7
Why does my AI character’s face keep changing?
Each generation starts fresh, so any change in the wording, the seed or the reference gives the model room to drift; small details slip first. Midjourney warns that intricate details like specific freckles or logos on clothing may not perfectly match your reference.8
Related on this site: ComfyUI Explained · First Frame and Last Frame in AI Video · What Is ControlNet? · What Is a Style Reference in AI? · Veo vs Kling vs Runway vs Seedance · Higgsfield Prop Sheets · How to Keep the Same Character in a Comic or Picture Book With AI · How to Train a LoRA for a Consistent Character
- 1Black Forest Labs, FLUX.2 Image Editing, read at source 13 Oct 2026“Reference multiple images simultaneously - up to 8 via API, up to 10 in the playground.” BFL’s full documentation file (docs.bfl.ai/llms-full.txt), read the same day, contains no mention of IP-Adapter. ↑
- 2OpenArt, OpenArt home page, read at source 13 Oct 2026“Create consistent AI characters across scenes” and “Create a character once, and you can place them in any outfit, scene, or video with the same face and look every time.” ↑
- 3Imagine Art, Imagine Art home page, read at source 13 Oct 2026“Over 50 image, video, audio and language models from providers including OpenAI, Google” and “Eight of the 50+ models you can switch between mid-workflow, without moving your work elsewhere.” ↑
- 4OpenAI Help Center, Images in ChatGPT, read in a browser 12 Oct 2026 and again by direct fetch 13 Oct 2026, when the page loaded without a browser“ChatGPT Images 2.5 brings sharper details, more precise editing, and faster image generation, alongside new ways to create and share.” ↑
- 5OpenAI Help Center, Images in ChatGPT, read in a browser 12 Oct 2026“To edit an image, select an image generated by ChatGPT Images or upload an existing image, then describe the change you want.” ↑a“To edit an image, select an image generated by ChatGPT Images or upload an existing image, then describe the change you want.” ↑b
- 6Google, Nano Banana image generation, read in a browser 13 Oct 2026Nano Banana 2.1 is “An update to Nano Banana 2, serving as the primary high-efficiency workhorse model for image generation and conversational editing.” ↑
- 7Google, Nano Banana image generation (Gemini API), read in a browser 12 Oct 2026“Nano Banana models allow you to mix up to 14 reference images.” The table lists “Up to 4 images of characters to maintain character consistency” for Gemini Nano Banana 2.1 & Gemini 3.1 Flash Image, “Up to 5 images of characters to maintain character consistency” for Gemini 3 Pro Image, and N/A for Gemini 3.1 Flash Lite Image. ↑a“Nano Banana models allow you to mix up to 14 reference images.” The table lists “Up to 4 images of characters to maintain character consistency” for Gemini Nano Banana 2.1 & Gemini 3.1 Flash Image, “Up to 5 images of characters to maintain character consistency” for Gemini 3 Pro Image, and N/A for Gemini 3.1 Flash Lite Image. ↑b
- 8Midjourney, Omni Reference, re-read directly at source 13 Oct 2026“Omni Reference is compatible with Midjourney version 7.” “You can only use one image with Omni Reference.” “You can set this parameter to any value between 1 and 1,000, with the default being --ow 100.” ↑a“Be aware that intricate details like specific freckles or logos on clothing may not perfectly match your reference.” ↑b
- 9Midjourney, Edit Model, re-read directly at source 13 Oct 2026“Generate new images using up to 4 reference images (replacing Omni Reference and Character Reference).” and “The Edit model is compatible with Midjourney versions 8.1 and 8.2.” ↑a“Shift the perspective and show different views of an object or character using the Edit model.” Example prompts: “let’s see this image from the front / let’s see this image from behind”. ↑b“If certain features or attributes are getting mixed up between your characters, it can be helpful to combine your characters into a single reference image.” ↑c“Generate new images using up to 4 reference images (replacing Omni Reference and Character Reference).” and “The Edit model is compatible with Midjourney versions 8.1 and 8.2.” ↑d
- 10Kling AI, Kling 4.0 vs 3.0, 30 Sep 2026, read at source 12 Oct 2026It lists “Up to 10 images and 7 subjects”. · Kling AI, Kling 4.0 at Work, 27 Oct 2026, read at source 3 Nov 2026: “Since Kling 4.0 rolled out to all users on October 15, creative teams across film, advertising, and animation have been putting these capabilities to work.” ↑
- 11Kling AI, Kling Element Library User Guide, 5 Feb 2026, read at source 12 Oct 2026each element can include “up to 4 reference images (1 main reference image + 3 supplementary reference images)”; element use is supported in “Kling VIDEO 3.0, VIDEO 3.0 Omni, Kling O1 Video and Image models”; in Video 3.0 Omni or Video O1 an input that includes a video allows “a total of 4 images/elements”, and up to 7 without one. ↑athe Kling O1 Image model can “automatically generate additional views from a single main reference image”. ↑bin multi-character scenes “the model can lock and maintain each character or prop’s features”. ↑c
- 12Google, Generate videos with Veo 3.1 in Gemini API, read at source 12 Oct 2026“Veo 3.1 now accepts up to 3 reference images to guide your generated video’s content. Provide images of a person, character, or product to preserve the subject’s appearance in the output video.” Note: “This feature is available for Veo 3.1 models only.” Duration: “Must be ‘8’ when using extension, reference images or with 1080p and 4k resolutions”. ↑a“Veo 3.1 now accepts up to 3 reference images to guide your generated video’s content. Provide images of a person, character, or product to preserve the subject’s appearance in the output video.” Note: “This feature is available for Veo 3.1 models only.” Duration: “Must be ‘8’ when using extension, reference images or with 1080p and 4k resolutions”. ↑b
- 13Dreamina (ByteDance), Seedance 2.5, read at source 12 Oct 2026“Supporting up to 50 multimodal inputs, Dreamina Seedance 2.5 understands complete creative direction.” Example prompt: “Use @Image 1 for the venue. Reference @Image 2 for the pianist.” ↑a“Supporting up to 50 multimodal inputs, Dreamina Seedance 2.5 understands complete creative direction.” Example prompt: “Use @Image 1 for the venue. Reference @Image 2 for the pianist.” ↑b
- 14Runway, Creating with Gen-4 Image References, re-read directly at source 13 Oct 2026“Use up to three References for a single generation”, “Tag your References to save them for future use”, and “Gen-4 References excels at generating consistent characters across different lighting conditions, locations, and treatments, all from just a single reference image.” ↑a“Use up to three References for a single generation”, “Tag your References to save them for future use”, and “Gen-4 References excels at generating consistent characters across different lighting conditions, locations, and treatments, all from just a single reference image.” ↑b
- 15Google, Nano Banana image generation, read at source 16 Sep 2026. An earlier version of this page named “Nano Banana Pro 2”; Google ships Nano Banana 2, Nano Banana 2 Lite and Nano Banana Pro.↑
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