ChatGPT Photo Editor: What a Chat Window Can't Do to Your Photo
Chat-based editing renders a brand-new image every turn. That one fact explains the drifting face, the garbled sign and the moved props — and it tells you exactly how to prompt around it.
August 13, 2026 · 7 min read
On this page
Quick answer
A GPT-class image model does not edit your photo the way a layer-based editor does. It reads the image, reads your instruction, and paints a new image that is supposed to look like the old one plus your change. Nothing is masked, nothing is locked, and nothing is preserved unless you say so in words. That single mechanic explains almost every complaint people have about editing photos in a chat box — the face shifts, the sign turns to gibberish, a prop moves across the table. You can work with it: name what must not change, make one change per run, and use an interface where ratio, resolution and reference images are controls rather than sentences you have to negotiate. On Shotari's GPT Image 2 editor an edit costs 5 credits; the built-in Shotari Basic model does the same round trip for 0.
Shotari is not ChatGPT
ChatGPT is OpenAI's product and this page is not it. What follows is about editing an existing photo with a GPT-class image model — the kind you reach from Shotari's model picker — and about what a chat transcript gives you versus a purpose-built editor. Every number quoted here is Shotari's own, read off the code that does the charging.
Why it changes what you didn't touch
In Photoshop, an edit is destructive only where you make it destructive. A selection, a mask, a layer — everything outside them is byte-for-byte the original file. A generative edit has none of that machinery. Your photo goes in as a reference, your sentence goes in as an instruction, and what comes back is a fresh render. The model is not moving your pixels around; it is drawing a new picture that it believes matches your description of the old one.
So the question is never 'why did it change that?' The question is 'why would it have kept that?' Anything you did not name is a detail the model is free to re-decide. Here is what that looks like in practice.
| What you ask for | What actually comes back |
|---|---|
| Crop this to 16:9 | A new 16:9 frame, with the edges re-imagined rather than kept. When you want the original pixels untouched and only the new area painted, that is a different operation — uncropping, not editing. |
| Remove the person on the left | They are gone and the fill is convincing. The rest of the frame has also drifted a little: grain, colour cast and small background objects are all redrawn. |
| Make the shirt red | A red shirt — and, sometimes, a face that reads a year older and a hair parting that has switched sides. |
| Keep the text on the sign | Small text is redrawn glyph by glyph. Long strings routinely come back as convincing lookalike gibberish. |
Clauses that pin what must not change
The fix is not a better one-line prompt. It is adding a second half to the prompt that says what to leave alone. These four clauses do most of the work, and they are the same clauses shipped inside Shotari's own templates for exactly this reason.
Identity lock
Preserve the person's exact face, hairstyle, skin tone and body shape. Change only the background.
Single-change lock
Change only the sky. Leave the subject, the foreground objects and the framing exactly as they are.
Product lock
Keep the product, its label text and its proportions identical. Rebuild only the surface it sits on and the lighting.
Text lock
Do not redraw any text in this image. Reproduce every word exactly as it appears.
Be honest about the ceiling here: a text lock reduces garbling, it does not remove it. If a phone number or a brand name has to be exactly right, generate the image without the text and add it afterwards in any editor that draws real type. More on writing instructions models actually follow: how to write AI image prompts.
The three-step edit
1. Give it the photo, not a description of the photo
Upload the original as a reference image rather than describing what is in it. Shotari's composer takes up to three reference images at once, which is what makes 'put this product on that background, in this style' a single run instead of three rounds of negotiation.
2. One change per run
Every run is a fresh render, so two instructions in one prompt compound the drift instead of adding up. Do the change, look at it, then run again from the result. On Shotari Basic each of those rounds costs 0 credits, so there is no reason to bundle them.
3. Hand the last mile to a tool that is not generative
Generative editing is the wrong instrument for some jobs. Cutting a subject out cleanly is background removal; widening a frame without touching the original pixels is uncropping; fixing a damaged scan is restoration. Each of those keeps more of your original than a repaint ever will.
What one edit costs
One run, one image, one charge. There is no per-message metering and no monthly quota to ration — you can see the price of a run before you press the button.
| Model | Credits per image | What it is for |
|---|---|---|
| Shotari Basic | 0 | Every experiment before the keeper |
| Shotari Pro | 2 | Higher quality; needs an account |
| Seedream 5 Lite | 3 | Cheapest premium model |
| GPT Image 2 | 5 | Prompt-driven edits, on-image text |
| GPT Image 1.5 | 5 | Faster edits, clean text |
| Nano Banana Pro | 10 | High-fidelity retouching |
Resolution and shape multiply that base: 2K is ×1.5, 4K is ×2.0, and a 9:16 or 16:9 frame adds ×1.1 on top, rounded up. GPT Image 2 at 4K in 16:9 is therefore 5 × 2.0 × 1.1 = 11 credits, not 5.
The part that surprises people: free credits and paid credits are not interchangeable. The 20 credits you get at signup and the 10 from each daily check-in are earmarked for Shotari's own models, so a balance of 30 will still refuse a 5-credit GPT Image 2 run. That is a deliberate rule, not a bug — the long version is in is GPT Image 2 free?.
Where it refuses, drifts or stalls
- ⚠
Most 'failures' are refusals, not crashes
When we split our own generation failures by path, text-to-image succeeded about 94% of the time and nearly every failure sat in the image-to-image path — and most of those were content refusals rather than technical faults: portraits the moderation layer reads as explicit, recognisable copyrighted characters, and anything shaped like an ID document. The tell is that the same prompt fails instantly on a different photo. See GPT Image 2 not working for the error-by-error breakdown.
- ⚠
Editing an image takes longer than generating one
A normal image-to-image round trip on our pipeline runs around 78 seconds, several times longer than a text-to-image run. A job that is 30 seconds in is not stuck — we learned this the expensive way, having once set a timeout shorter than the job it was timing.
- ⚠
The free download carries a watermark
Guests can generate but not download; the download button asks for an account first. A signed-in free account downloads with a shotari.com watermark, and the watermark-free file needs paid credits. Nothing about that changes with the model you pick.
- ⚠
Without an account you get two runs a day
Two generations per network per day on Shotari Basic, counted on a UTC day boundary. Shotari Pro requires signing in before the first run, not after it.
FAQ
Can ChatGPT edit an existing photo?
Yes — chat-based image models accept an uploaded photo and hand back an edited version. What they do not do is edit in place. The reply is a fresh render of the whole picture, which is why details you never mentioned come back changed.
Is there a free ChatGPT photo editor?
Free editing yes, free GPT Image 2 no. Shotari Basic runs at 0 credits and the first two runs need no account at all, but GPT Image 2 specifically costs 5 paid credits per image. Anyone promising a premium model for nothing is either time-limiting it or watermarking it.
Why does the face change when I edit my photo?
Because the model redraws the person along with everything else. Add an explicit clause — 'preserve the person's exact face and identity' — and it holds; leave it out and the face is just another detail the model gets to re-decide.
How do I edit a photo online without an account?
Open the editor, upload your photo and run it on Shotari Basic. You get two runs per day without signing in. Downloading the result is where an account becomes mandatory.
Can I upload more than one photo?
Up to three reference images per run. That is what lets you combine a subject, a background and a style reference in a single instruction instead of stacking three separate edits.
Which model should I use for edits with text in them?
GPT Image 2 or GPT Image 1.5, both at 5 credits — on-image text is the thing they are best at. Draft the composition on Shotari Basic for free first, then spend the 5 credits once on the version you intend to keep.
Edit the photo, not the transcript
Upload your image, name what must not change, and run it. Free rounds on Shotari Basic, 5 credits when you switch to GPT Image 2 for the keeper.
Open the GPT Image 2 editor