OpenAI’s ChatGPT Images update finally addresses the complaint that has dogged the product since day one: ask for one small change to a photo of a person, and the model hands you back a stranger who vaguely resembles them. The new release targets identity persistence across edits for human faces and pets, and adds genuinely localized editing, so a request to change the lighting on a wall no longer rerenders the subject’s entire face. If you use ChatGPT for product shots, family photos, thumbnails, or client mockups, this is the first version where an iterative workflow — edit, review, edit again — holds together. It also means the defensive prompting habits most people built over the past two years now work against them.
What’s actually new in the ChatGPT Images update
The core change is identity anchoring. Previous versions treated every edit as a fresh generation conditioned loosely on the source image, which is why faces drifted. Each pass re-derived the subject’s features from a text description plus a fuzzy visual reference. The updated model maintains a persistent identity representation across a conversation, so the same face — or the same dog, with the same asymmetric ear and the same white chest patch — survives four or five sequential edits instead of degrading after one. OpenAI frames this as ChatGPT face consistency, and it shows up most clearly in multi-turn sessions rather than single-shot generations.
The second change is localized editing, closer to true inpainting than the old whole-frame regeneration. Ask for a change confined to a region — remove the coffee cup, change the shirt from blue to charcoal, brighten the background — and the model edits that region while leaving the rest of the pixels substantially intact. Previously, every edit was a full redraw. Backgrounds shifted, text in signage garbled itself, and hands that were fine in pass one became unusable in pass two. ChatGPT inpainting edits are now selective enough that you can iterate on a composition without treating each pass as a gamble.
Pets got explicit attention, which sounds like a novelty feature until you consider volume. Pet photos are one of the highest-frequency consumer use cases, and animals were harder than humans because breed markings, fur texture, and eye color all vary in ways the model previously smoothed toward a generic average. The ChatGPT pet photos AI improvements mean specific markings persist, not just species and general coloring. Multi-subject scenes — two people, or a person and their dog — hold identity better too, though this remains the weakest area: the more subjects in frame, the more likely one drifts.
Why it matters
- Iterative workflows become viable. The old model forced a one-shot mentality — get it right in a single prompt or start over. You can now treat generation as a conversation, which is how designers actually work.
- Your old prompts are now counterproductive. Everyone learned to over-describe subjects (“a woman, 30s, brown eyes, shoulder-length dark hair, oval face”) to fight drift. Those descriptions compete with the model’s identity anchor and can pull the face away from the source.
- Small-business production costs drop. Product photography, real estate listing cleanup, and social thumbnails were bottlenecked by the redraw problem. Localized edits mean one good base image can generate a dozen usable variants.
- The gap to Midjourney narrows on control, not aesthetics. Midjourney still wins on raw image quality for stylized work. ChatGPT now wins on “change this one thing and don’t break everything else,” which matters more for commercial output.
- Consent and likeness questions get sharper. A model that reliably preserves a specific person’s face across edits is more useful and more abusable. Expect platform policy and detection tooling to tighten in response.
- Reference-image workflows finally pay off. Uploading a source photo used to be a suggestion; it now functions closer to a constraint, which changes what’s worth uploading.
How to use the ChatGPT image editing 2026 workflow today
-
Start with the source image and a minimal prompt. Upload the photo, then describe only the change. Resist describing the subject. This is the single biggest adjustment for experienced users.
[upload photo] Change the background to a plain warm gray studio backdrop. Keep the subject exactly as-is — same face, same hair, same pose, same lighting on the skin. -
Scope edits explicitly by region. Naming what should not change now works better than it used to, because the model has a mechanism to honor it.
Edit only the region of the countertop. Remove the two coffee cups and the newspaper. Do not alter the person, the window, or the lighting anywhere else in the frame. -
Chain edits in one conversation, not across new chats. The identity anchor lives in the session. Starting a fresh chat and re-uploading the output of a previous edit throws away the anchor and reintroduces drift. Do all passes in a single thread.
-
For pets, name the distinguishing marks once, on the first turn. After that, stop repeating them.
[upload photo of dog] This is Rosie — a brindle boxer mix with a white blaze down her muzzle, one folded left ear, and a white left front paw. Keep these exact. Now: place her sitting on a hardwood floor in soft window light. -
Use the API for batch or programmatic work. The Images endpoint accepts the source image plus a prompt. Pass the same source across variants rather than feeding output back in as input, which compounds artifacts.
curl https://api.openai.com/v1/images/edits \ -H "Authorization: Bearer $OPENAI_API_KEY" \ -F model="gpt-image-1" \ -F image="@subject.png" \ -F prompt="Replace the background with a neutral gray studio backdrop. Preserve the subject unchanged." \ -F size="1024x1024" \ -F n=1 -
Add a mask when you need hard boundaries. Prompt-level scoping is good; an explicit alpha mask is deterministic. The transparent region of the mask is what gets regenerated.
curl https://api.openai.com/v1/images/edits \ -H "Authorization: Bearer $OPENAI_API_KEY" \ -F model="gpt-image-1" \ -F image="@subject.png" \ -F mask="@background-mask.png" \ -F prompt="Warm afternoon light through a window, soft shadows." \ -F size="1024x1024" -
Verify at full resolution before you ship. Drift now shows up in small details rather than whole faces — teeth, jewelry, logo text, the shape of a nostril. Download the output and view it at 100 percent. Thumbnail review will pass work that fails on a client’s monitor.
-
Keep a clean base. Save the best generation as your canonical source and branch from it. Never build a chain more than three or four edits deep from an already-edited file; go back to the base instead.
How it compares
| Capability | ChatGPT Images (updated) | Midjourney | Google Gemini / Imagen | Adobe Firefly |
|---|---|---|---|---|
| Face consistency across edits | Strong — persistent identity anchor in-session | Moderate — character reference helps but drifts | Moderate to strong, improving quickly | Moderate; strongest with Photoshop layers |
| Localized editing / inpainting | Native, prompt-scoped or masked | Region editor, requires manual selection | Prompt-based, less precise boundaries | Best-in-class via Generative Fill |
| Pet and animal fidelity | Explicitly improved this release | Good aesthetics, weak on specific markings | Good | Adequate |
| Conversational iteration | Native — the core strength | Discord/web commands, less fluid | Native in Gemini app | Panel-based, not conversational |
| Text rendering in images | Reliable for short strings | Improved but inconsistent | Reliable | Reliable |
| API availability | Yes, images endpoint | No official API | Yes | Yes, Firefly Services |
| Best for | Iterative commercial edits | Stylized art direction | Speed and integration | Professional retouching pipelines |
What’s next
The obvious next step is persistence across sessions. The identity anchor is scoped to a conversation, which limits anyone building a repeatable brand asset — a mascot, a recurring spokesperson, a consistent character for a series of posts. Saved subject profiles, where you register a face or pet once and reference it from any chat, would turn this from a nice improvement into a production tool. Watch for it, and watch how OpenAI gates it, because a saved-likeness feature carries consent implications that a per-session anchor does not.
Video is the other direction. Identity consistency is the same fundamental problem in Sora and every other video model, just harder — the face has to hold across hundreds of frames, not four edits. Improvements to the image model’s identity representation tend to migrate into the video stack, so this OpenAI image model update signals where Sora goes next. If your work touches short-form video, the practical question is when you can generate a character in ChatGPT Images and carry that exact character into a clip.
The near-term thing to watch is how the competition responds. Google ships image improvements on a fast cadence and holds distribution advantages through Android and Workspace. Adobe owns the professional retouching workflow and will position localized AI edits as a Photoshop feature rather than a chat feature. The interesting scenario is not one winner but specialization: chat interfaces for fast iterative work, professional tools for final output, with a handoff between them. For now, the most useful GPT image prompt tips remain the counterintuitive ones — say less about the subject, be specific about the region, and stay in one thread.
Frequently Asked Questions
Do I need to change how I write prompts?
Yes, and this is the main practical takeaway. Detailed subject descriptions that used to fight drift now compete with the model’s identity anchor. Describe the change you want, name what should stay fixed, and let the source image carry the subject’s appearance.
Does face consistency work across separate chats?
Not reliably. The identity representation is scoped to the conversation, so re-uploading an edited image in a new chat starts over and reintroduces drift. Do all your passes in a single thread, and keep your best generation as a canonical base to branch from.
How many edits can I chain before quality degrades?
Three or four sequential edits hold up well; beyond that, small details accumulate errors even when the overall face is preserved. Rather than building a long chain, go back to your base image and apply a combined edit.
Is this available on the free tier?
Image generation limits differ by plan, and OpenAI typically rolls model upgrades to paid tiers first before broadening access. Free users generally get the newer model with tighter daily caps. Check your plan’s current image limits in ChatGPT settings rather than assuming.
Can I use this for editing photos of real people commercially?
Technically yes; legally it depends on consent and jurisdiction. You need permission from anyone whose likeness you edit and publish, and that requirement gets stricter, not looser, as the tool gets better at preserving identity. Get releases in writing for client work.
How does it handle multiple people in one image?
Better than before, but this is still the weakest area. With two subjects you’ll usually keep both; with four or five, expect at least one to drift. If a group shot matters, edit subjects in separate passes on isolated regions rather than asking for a change that touches everyone at once.
Go deeper than this article
This article covers the essentials. Our premium eguide library gives you the full step-by-step playbooks — prompts, workflows, and copy-paste recipes you can put to work today.
