ChatGPT Images 2.6 Reference Photos: Does It Keep Your Face?

ChatGPT Images 2.6 can save a photo of you and reuse it in any chat. What that fixes, what it doesn't, and a 5-minute test to check your likeness.

GetPhotoShoot Team··7 min read
A phone showing a saved ChatGPT Images 2.6 reference photo next to a row of generated portraits where the face shifts slightly in each one

ChatGPT can now remember a photo of you. With ChatGPT Images 2.6, released by OpenAI on November 4, 2026, you save a reference image once, name it, and pull it into any later chat with an @ mention. No re-uploading, no hunting through your camera roll.

That fixes a real annoyance. It does not fix the reason your ChatGPT headshot never quite looked like you. A saved photo is still one photo, and the model still reads it the same way it did in Images 2.0. Below is what changed, what didn't, and a five-minute test that tells you whether the result is safe to use on LinkedIn or a dating app.

Your photos, any style

See it on your own photos

Upload a few selfies and pick a style. You pay once, with no subscription.

Drop your photos here

or click to browse · JPG, PNG

One-time · no subscriptionWorks with phone photosDeleted after 30 daysCredits never expire

What ChatGPT Images 2.6 Actually Added

Four things shipped in the OpenAI announcement:

  1. Reference images. Save pictures of yourself, a pet or a character and reuse them in new images without uploading again. A saved image gets a name you can call with an @ mention in later conversations.
  2. GIFs. Turn a prompt or an uploaded image into a looping animation, adjust the speed and edit single frames.
  3. Speed. OpenAI says generation is 50% faster. On the API, the new Flare model can return an image in about six seconds at low quality.
  4. Two API models. Flare for speed, Sunburst for quality and for keeping details consistent across a chain of edits.

Availability is uneven. Coverage of the launch describes 2.6 as live across subscription tiers while GIFs and named references are still rolling out. If you do not see the save option yet, nothing is wrong with your account.

Stat

Images 2.6 is the third ChatGPT image update since April 2026: Images 2.0 in April, Images 2.5 with virtual try-on in October, and 2.6 on November 4.

Saved References Fix Friction, Not Likeness

Here is the distinction most launch coverage skips. A reference image is input. The model looks at it while it draws. It does not learn from it, and it does not keep anything about your face between generations except the file itself.

So what does a saved reference really change?

  • Before: you uploaded a selfie into each new chat, often a different selfie each time. Every headshot drifted from a different starting point.
  • After: every generation starts from the same file. Your images drift less from each other.

That is consistency. It is not accuracy. If your saved photo has soft window light that hides your jawline, every image built from it invents a jawline. If you are smiling in it, the model has to guess what your face does at rest. It guesses the same way more often now, which can make a wrong guess look more convincing, not less.

We covered the root cause in detail in our Images 2.0 headshot breakdown. One photo is one angle, one light and one expression. Everything outside that frame is generated.

Where "Consistency Across Edits" Helps and Where It Doesn't

OpenAI pitches Sunburst as better at keeping details steady across edits. That claim is about a chain: you make an image, then ask for "same photo, navy blazer," then "same photo, grey backdrop." Each step changes a little and keeps the rest.

That is useful for headshots. Change the shirt without the face sliding around is a real improvement over April.

It is a different job from putting you in a new scene. A café dating photo, a hiking shot and an office portrait are three fresh generations. Each one rebuilds your face at a new angle, in new light, from the same flat reference. That is where people see the drift, and edit consistency does not reach it.

Even OpenAI's own users noticed. In the replies under the announcement, one developer reported that a 2.6 test portrait was "no longer identifiable" as the person in the source photo.

The 5-Minute Likeness Test

Before you put a ChatGPT Images 2.6 portrait anywhere a stranger will compare it to your real face, run this:

  1. Save your best reference photo: even light, neutral expression, face straight on.
  2. Ask for six images in clearly different settings. Office, café, outdoors at golden hour, a studio backdrop, a three-quarter turn, a full-body shot.
  3. Put all six in a row next to the original photo.
  4. Check four features only: eye spacing, hairline, nose width and jawline. Skin and hair texture will always look good. Those four are where likeness breaks.
  5. Show the row to one friend and ask which images do not look like you. Do not tell them which is real.

If your friend flags two or more of the six, the reference is doing its job but the method is hitting its limit. More prompting will not change that. We explain why in why AI headshots look different every time.

Get headshots that pass the likeness test →

Upload a few selfies, we train on your face, results in minutes.

Saved Reference vs a Model Trained on Your Photos

ChatGPT Images 2.6 referenceModel trained on your photos
Photos it sees110 to 20
Angles and expressions it knowsThe ones in that photoMany, from your uploads
What happens in a new sceneFace rebuilt from one viewFace drawn from learned features
Best atDrafts, edits, outfit and setting ideasProfile photos that must match you
Setup timeSecondsMinutes

Neither column is "better" in general. They solve different problems. If you need ideas for what to wear, where to stand and which background suits you, Images 2.6 is fast and cheap, and the GIF feature is fun. If you need a photo for your LinkedIn profile or for a dating app, the person on the other side will meet you or call you. That is the case where likeness is the whole job.

The Privacy Setting to Check Before You Save Yourself

A saved reference photo of your face lives on your account by design. That is the point of the feature. Two things are worth doing before you save one:

  1. Know where to delete it. The Images 2.6 announcement does not say where reference images are managed. The reference photo from ChatGPT virtual try-on, launched in October, sits under Settings → Personalization → Reference photos. Check there first.
  2. Check model training. On personal plans, uploads can be used to improve OpenAI's models while Improve the model for everyone is switched on in Data controls. Turn it off if you do not want your face in that pool.

When Images 2.6 Is the Right Tool

Use it when the image is for you, not for someone verifying you:

  • Testing outfits and colors before a real shoot or before you order a pack
  • Mood boards for a photographer
  • Fun content: GIFs, pet portraits, a recurring character
  • Editing an existing photo in small steps, where Sunburst-style consistency shines

Skip it for anything a recruiter, client or match will hold up against your face on a video call. If your ChatGPT portrait already doesn't look like you, a saved reference will make the same mismatch repeat more reliably.

The short version: Images 2.6 makes ChatGPT a better sketchbook. For the final photo, train on your face.

Train an AI on your face, not on one photo →

Pick LinkedIn, dating or a full photo session. Done in minutes.

Frequently asked questions

What is new in ChatGPT Images 2.6?

OpenAI released ChatGPT Images 2.6 on November 4, 2026. It adds saved reference images of yourself, your pets or recurring characters, looping GIF creation from a prompt or an upload, and generation that OpenAI says is 50% faster. Developers got two API models, Flare for speed and Sunburst for quality and consistency across edits.

How do ChatGPT reference images work?

You save a photo once and give it a name. In a later conversation you call it up with an @ mention instead of uploading the file again. The model then uses that photo as a visual reference for the new image. It does not learn your face from it. It reads the same single photo every time.

Does a saved reference photo make ChatGPT headshots look like me?

It makes them more consistent with each other, because every generation starts from the same file. It does not make them more accurate. One photo gives the model one angle, one light source and one expression, so any feature hidden or flattened in that photo still gets guessed, and the guess still shifts between scenes.

Is ChatGPT Images 2.6 free?

OpenAI says Images 2.6 is available across ChatGPT subscription tiers, but GIF creation and named reference images are still rolling out, so some accounts do not see them yet. Image limits follow your plan, so free accounts hit the ceiling sooner than paid ones.

Where are my ChatGPT reference photos stored?

They stay attached to your account so you can reuse them. OpenAI's Images 2.6 announcement does not say where you manage them. The reference photo used by ChatGPT virtual try-on, launched in October 2026, sits under Settings, Personalization, Reference photos, which is the first place to check and delete one.

Should I use ChatGPT Images 2.6 for a LinkedIn or dating profile photo?

Use it for drafts and for ideas about outfit, background and light. For a photo that strangers will compare against your real face, use a tool that trains a model on 10 to 20 of your own photos. A trained model holds jawline, hairline and eye spacing steady across every scene.

Try it free

Transform your photos with AI

Upload a selfie and get stunning AI-generated photos in seconds. Free preview — no credit card required.

Start Taking AI Pictures

No credit card required. Results in 60 seconds.