ChatGPT Consistent Character: Keeping Faces Stable in GPT Image Generations

To get a consistent character image in ChatGPT, approve one hero image, upload it back as a reference at the start of every session, paste the same character description block, and ask for one change at a time while listing what must stay the same. OpenAI's help pages describe no saved-character setting for ChatGPT Images, so the method is to hand it the identity again each time: the picture and the words.

Based on OpenAI's documentation, checked October 2026: the Images in ChatGPT, Projects, file uploads and Free tier help articles, the GPT Image 2.5 prompting guide, the image generation guide, the deprecations page and the usage policies. The templates below follow those pages. They are patterns to adapt, not prompts we scored.

This guide is for anyone building an original AI persona, or a recurring version of themselves, inside ChatGPT rather than a dedicated image tool. It covers what ChatGPT uses for images now, the description block to re-paste each session, the edit prompts that protect the face, and the point where ChatGPT stops being the right tool. For how this method compares with reference models, ID adapters and LoRAs, read our consistent character AI guide first.

DALL-E consistent character advice is out of date: what ChatGPT uses now

Most tutorials for this topic were written for DALL-E 3. That model is gone. OpenAI's help center says the official DALL-E GPT has been retired from ChatGPT, and image requests now run on ChatGPT Images 2.5, which the same page describes as bringing sharper details, more precise editing and faster generation.

Two things follow for a persona. First, old tricks tied to DALL-E, such as asking for a generation ID or a seed, have no equivalent in the ChatGPT Images help page or the API guide. Second, the new models are better suited to the job: OpenAI says both GPT Image 2.5 models improve precise editing and subject preservation. It also says, in the limitations list of the image generation guide, that the model "may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations". Plan for both statements to be true.

The system: how a ChatGPT persona holds together

A consistent character in ChatGPT is four assets and one habit. The assets are a description block, a hero image, a small set of reference views and a place to keep them. The habit is restating identity in every request.

Nothing here depends on ChatGPT remembering you. That is deliberate: what carries over between chats is not something you control, and a method that leans on it fails quietly.

The description block to re-paste each session

The block has three labelled parts: the character, which never changes; this image, which is the only part you edit; and a keep line that restates identity last. OpenAI's prompting guide recommends labelled sections for complex requests and says to separate changes from constraints, so the layout follows the documentation rather than a private trick.

CHARACTER (never changes):
[Name], a fictional [age]-year-old [woman / man].
Face: [face shape], [eye colour and shape], [brows], [nose], [lips].
Hair: [colour], [length], [texture], [parting].
Marks: [one or two fixed details].
Build: [height impression], [build].
Reference: the attached image is [Name]. Match the face, hair and proportions exactly.

THIS IMAGE (the only part I change):
Scene: [place], [time of day], [light].
Action and framing: [pose], [gaze], [crop].
Outfit: [clothing for this image].
Look: real photograph, [lens or feel], natural skin texture.

KEEP: same facial features, skin tone, hair and proportions as the reference. No text, logos or watermarks.

Filled in for an illustrative persona:

CHARACTER (never changes):
Mara, a fictional 28-year-old woman.
Face: oval face, wide-set hazel eyes, straight dark brows, narrow nose, full lower lip.
Hair: dark brown, collarbone length, loose waves, centre parting.
Marks: a small mole under the left eye, light freckles across the nose.
Build: medium height, slim build.
Reference: the attached image is Mara. Match the face, hair and proportions exactly.

THIS IMAGE (the only part I change):
Scene: a small cafe by a window, late morning, soft overcast light.
Action and framing: seated, stirring a coffee, looking at the camera, waist-up.
Outfit: cream knit sweater, thin gold chain.
Look: real photograph, 50mm feel, shallow depth of field, natural skin texture.

KEEP: same facial features, skin tone, hair and proportions as the reference. No text, logos or watermarks.

• Describe what a camera would see. Bone structure, eye shape, hair, fixed marks. Leave out judgement words such as beautiful or stunning; they pull every face toward the same average.
• Keep clothing out of the character part. A character who is "wearing a white dress" fights every other outfit you ask for.
• Freeze the wording. Store the block in a text file and paste it. A synonym is a different instruction.
• State that the person is fictional and adult. It is true, and it keeps the request unambiguous.

If you would rather generate the character part than write it, our free AI influencer prompt generator builds an identity block from a persona brief, and the AI influencer prompt pack has 100 scene lines that slot into the "this image" part.

ChatGPT same character in multiple images: the session workflow

Run every session the same way. The first two steps happen once; the rest repeat each time you sit down to make images.

In the app, you can ask for an image in any conversation, or select More and then Images. Every image you create is saved under Images, which is where you find the hero again later. For aspect ratio, use the picker or name the ratio in the prompt; a persona feed usually wants one ratio throughout.

Edit prompts that protect the face

An edit prompt has two halves: the one change, and the list of things that must not move. The patterns below are adapted from the examples in OpenAI's prompting guide. Swap the bracketed parts for your own.

The selection tool in the editor lets you highlight an area before describing the edit. OpenAI notes that highlights are not always precise and edits may extend beyond the area you selected, so an outfit edit can still touch the jawline. Check the face after local edits too.

Where to keep the block and the references

Use a Project. OpenAI's help page says Projects keep related chats, files and instructions together, that you can add images as project files, and that project instructions apply only within the project and override your global custom instructions.

• Project instructions: paste the character part of the block there, so every chat in the project starts with it.
• Project files: add the hero and the reference views so they are in one place. The number of files depends on your plan.
• Still attach the hero in chat. The help page describes project files as context for answers; it does not say image generation reads them as visual references. Attaching the image in the message removes the doubt.

A custom GPT used to be the usual answer here. OpenAI's Free tier FAQ now says new GPT creation and publishing are not available on personal accounts, including Free, Go, Plus and Pro, so a Project is the route for an individual creator.

Consistent character GPT free: what the Free tier allows

You can run this workflow without paying, slowly. OpenAI says ChatGPT Images is available on all tiers and lists creating images and uploading images among Free features, each with its own usage limit.

• Uploads: the file-uploads page limits Free users to three file uploads a day, against 80 files every three hours on paid plans, with a 20 MB cap per image. Three uploads is one hero and two references, or three sessions with the hero alone.
• Generations: no number is published. ChatGPT notifies you when you reach the limit.
• Images with thinking: listed for Plus, Pro and Business, not Free.

On Free, spend the uploads on the hero and make every message count: one change, clearly stated. If you hit limits daily, that is the signal to upgrade or move the persona to a tool built for volume.

The same method through the API

If you automate posts, the workflow carries over unchanged. The image edits endpoint accepts one or more reference images, and the Responses API supports multi-turn editing, so the hero travels with each request and the block is a fixed string in your code.

• Models: gpt-image-2.5-sunburst where editing precision matters most, gpt-image-2.5-flare for fast everyday generation.
• Quality: low to max. OpenAI suggests low for drafts and comparing higher settings for finals; a higher setting is not always a better result.
• Price: both models bill $8 per million image input tokens and $30 per million image output tokens (checked October 2026). Reference images count as input, so a three-reference request costs more than a text-only one.
• Access: OpenAI says you may need to complete API Organization Verification before using GPT Image models.

Where ChatGPT runs out for a persona

ChatGPT is the easiest place to start a character and a hard place to run one at volume. The help page lists no strength control for a reference and no seed field, and you cannot train the image model on your own pictures. When you find yourself rejecting more images than you keep, the fix is a different tool, not a longer prompt.

• More reference slots: our ranking of consistent character methods without a LoRA covers models that take several character images at once, and the Gemini face consistency prompts give you 30 templates for Nano Banana.
• Diagnosing drift: the AI face consistency guide lists the causes one by one.
• A face stored in weights: the LoRA training guide is the long-run answer for a persona you post daily.

The images you approve in ChatGPT are not wasted when you move. They become the reference set, and later the dataset, for the next method. The AI Influencers program goes deeper where this guide stops, with lessons on creating the persona, prompt engineering, LoRA training for consistent faces and photorealism.

Whose face you can use

Use an original persona or yourself. OpenAI's usage policies prohibit using someone's likeness, including their photorealistic image, without their consent in ways that could confuse authenticity, and they prohibit impersonation. That rules out building a "consistent character" from a celebrity, a creator you admire or a stranger's photos.

Check your hero image for accidental resemblance before you commit a feed to it; our AI likeness rights guide explains the risk. When the persona goes live, label it where platforms require: see Instagram's AI label rules.

ChatGPT consistent character: FAQ