Our AI image editing test pushes Tsubaki.3 from one simple edit to four at once, showing which details survived and which quietly moved.
In traditional art, a correction stays where you put it. When an animator repaints one cel, the rest of the scene stays exactly where it was, because the artist decides how far the change reaches.
AI image editors enable something similar in a sentence. Upload a picture, describe what you want changed, and the model handles it. But it can be tricky to control for beginners.
Suppose you have an image of a character at a café table, and you ask for her jacket in red. It comes back red, and her face has quietly changed shape. Or, you ask for the coffee cup to be removed, and the table redraws itself underneath. You try again and ask for a smile, and one earring is gone.
The instruction was followed every time, so nothing looks obviously broken. And yet what you got back isn't exactly the image you wanted.
In each version, the requested change was made, and something else quietly changed alongside it.
That gap is the real test of an AI image editor. Making the requested change is only half the job. The other half is leaving everything you never mentioned exactly as it was, and that part rarely gets tested, because an edited image tends to be judged on whether it looks good rather than whether it is still the same image.
So we ran an AI image editing stress test in PixAI, an anime-focused AI art generator. Every edit here was made with Tsubaki.3, PixAI's latest image model, built around control and consistency rather than one-off illustrations. It can be used to edit an existing image from a plain-language instruction while holding the rest of the picture steady.
We started with one simple change and kept raising the difficulty until preservation began to slip. Meet Rei, the character we tested on. She is a club DJ in her late twenties, and she came out of a single generation with a handful of details we can count: a platinum streak over her right ear, a lightning bolt patch on her right sleeve, silver headphones, and fingerless gloves on both hands.
Rei, generated in one pass with Tsubaki.3, the single image every edit in this article was made from.
Every one of those details is either on one specific side or countable, which is deliberate. Vague traits can't be graded. A streak either stays on her right or it doesn't.
Over the next few sections, we'll make one edit, then two, then three at once, and then push it further with a change to her outfit, her surroundings, and the light all together. Let's get started!
What Are We Testing in This AI Image Editing Test?
Before running anything, let's define what counts as a successful edit, because "it looks good" isn't something you can actually check. Every test below gets graded on two things separately.
The first is instruction following. Did the Tsubaki.3 model do what was asked? If the prompt contained three changes, did all three happen, or did one get dropped or only partly applied?
The second is preservation. Did everything else stay put? That covers Rei herself, including her face, hair, and proportions. It also covers her pose, the composition, the art style, the lighting, and every object around her that was never mentioned.
Those two factors come apart more often than you might expect, which is exactly why we are grading them separately. An edit can follow the instruction perfectly and still hand back a different person. It can also preserve everything beautifully and simply fail to make the change you asked for. That failure is easy to spot. The other one isn't, because nothing in the result tells you what used to be there.
So preservation has to be defined before it can be graded. Whether an AI image editor preserves details you never mentioned depends on which details you named as worth tracking.
Before the first edit, we marked the details in Rei's image that either sit on one specific side or can be counted.
The details we tracked through every edit are marked on Rei's original image.
Also, there's one setup detail to keep in mind. When you edit an image with Tsubaki.3, put the image you want to change in the first reference slot, and set the aspect ratio to "Auto". Every test below uses that same setup.
The editing setup, with Rei's original image in the first reference slot and the aspect ratio set to "Auto".
One more rule stayed the same throughout our AI anime image editing process. Every edit starts from the same original image, not the previous result. That way, each test measures one change against a fixed baseline instead of accumulating whatever the last edit left behind.
Test 1: Simple Targeted AI Image Editing
We started with two deliberately easy edits. When the instruction leaves almost nothing to interpret, whatever changes is the model's own doing.
The first prompt was five words as follows:
"Change her eyes to blue"
This was our result.
Both irises changed to blue while the eyeliner, lash detail, and brow shape stayed exactly as they were.
Tsubaki.3 followed the instruction cleanly. Both irises went blue, and the edit stopped there, which is the part to watch, because the easiest way to fail this test is to redraw the whole eye and hand back a slightly different face.
Preservation held across the whole checklist. The platinum streak stayed on her right, the three hoops stayed three, and the lightning bolt patch kept its shape and its position on the sleeve. Both gloved hands sat exactly where they were on the decks, and the ceiling, the two downlights, and the pale doorway behind her didn't move.
The second edit asked for a removal instead of a color change, which is the harder job. Recoloring only asks the model to repaint what is already there, while removing something means inventing whatever was hidden behind it.
We used this prompt:
"Remove the headphones from around her neck"
And we got this output.
The headphones and their cable are gone, and the model rebuilt a ribbed collar underneath that was never visible in the original image.
The headphones were removed, and the cable running down from them went too. That is the right call, since the cable belonged to the object we asked it to take out.
What replaced them is the interesting part. The original never showed her jacket collar underneath, because the headphones were sitting on it. Rather than leaving a gap or smoothing the area flat, the model built a ribbed collar matching the ribbing at her cuffs and hem. Nothing in the prompt asked for that. It's a reasonable guess about a part of the garment the model had only ever seen partly covered.
Preservation held again. Her face and hair came back unchanged, including on the side where the headphone cup had been sitting. The platinum streak is still on her right, the three hoops are still three, the patch is still on her sleeve, and both gloved hands are still on the decks.
Test 2: Structural AI Anime Image Editing That Rebuilds Part of the Frame
A colour swap asks the model to repaint something. But a structural edit asks it to rebuild something, and that is more complicated.
When an arm has to move, a hand has to close around an object, or a garment has to be redrawn from scratch, it leaves more room for errors. The result to look out for isn't just whether the edit lands, but how much of the surrounding picture survives the reconstruction.
We ran these edits, each starting from the same original image. The first changed her pose using this prompt:
"She is holding one hand up to her headphones and looking to her left"
This is how the edit turned out.
Her right arm lifts and her hand closes around the headphone cup, and the headphones move from her neck onto her head to make the gesture work.
The gesture landed, and the anatomy holds up, which matters because a raised arm is one of the easier ways to get a broken elbow or a hand with the wrong number of fingers. The patch traveled with the sleeve rather than staying pinned to the same spot on the canvas, which is correct.
The gaze moved, but the wrong way. We asked her to look to her left and got her right, which is the sort of mix-up you get when a prompt says left without saying whose left it means.
One thing came back unasked for. The headphones moved from her neck onto her head, which the prompt never requested. We asked for a hand raised to her headphones and left the rest open, so the model filled the gap with the most common version of that pose.
The second edit added an object she had to hold. We used the following prompt:
"She is holding a vinyl record up in her right hand"
The output we got is shown below.
A record appears in her raised hand with her fingers curled around the edge, while her other hand stays on the decks.
This was the cleanest of the three. Her fingers grip the record, her left hand never moved off the mixer, and the patch, streak, hoops, and headphones all stayed the same.
Meanwhile, the third edit changed her outfit entirely. This was the prompt we used:
"Replace her black bomber jacket with a cream cable-knit cardigan"
And the result came out like this.
The cardigan replaced the jacket cleanly, and the lightning bolt patch left with it.
Composition preservation was strongest here, and nothing moved. It's the same pose, same hands, same framing, same room, and same light.
But the lightning bolt patch is gone. It was one of the details we marked before starting, and the moment the jacket left, the patch left with it.
The model treats the patch as part of the jacket rather than part of Rei. The headphones sit on her, so they survived the same edit without being mentioned. The patch sits on her jacket, so it left when the jacket did. Any detail attached to a garment is read as part of that garment, which means you have to name it when you replace the thing it sits on.
Here's what that looks like:
"Replace her black bomber jacket with a cream cable-knit cardigan. Keep the single yellow lightning bolt patch on one sleeve only, in the same position it appears in the original image."
And here's the output.
The cardigan replaced the jacket, and a single lightning bolt patch came back on the same sleeve it started on.
Test 3: Complex AI Image Editing With Two or Three Changes at Once
Everything so far has been a single instruction. Real edits are usually messier.
You look at an image, notice three things you want to change, and ask for all three at once. The model then has to make each edit without letting one interfere with the others.
To test this, we ran the same edit twice. The first prompt shown below included a preservation clause. The second didn't.
"Change her hair to a long ponytail, replace her black bomber jacket with a white tank top, and change her expression to a wide open-mouth grin. Keep the single platinum streak at the front on her right, keep the three silver hoops in her left ear, and keep the silver headphones at her neck."
Then we removed the second sentence and ran it again.
The same three changes run twice, with the preservation clause in the middle and without it on the right, where a second platinum streak runs the length of the ponytail.
Both runs made all three changes. The hair came up into a long ponytail, the bomber jacket became a white tank top, and the expression became a wide open-mouth grin. Nothing was dropped or partly applied either, which isn't a given when three instructions arrive in one sentence.
The ponytail was the change most likely to cause trouble. Pulling her hair back exposes the ear where the three hoops sit and clears the shoulders that the tank top then had to render, so two of the three edits were working on the same part of the frame. Both runs handled it.
The difference is in the streak. The run with the clause kept exactly one platinum streak, at the front on her right, where it started. The run without the clause kept that one and added a second, running the full length of the ponytail.
That is the same pattern as the lightning bolt patch on the cardigan. Left to describe a trait rather than reproduce it, the model treats it as a feature of the category rather than one object in one place. Saying "the single platinum streak at the front on her right" pinned down both the count and the location, and that held.
The rest of the checklist came through in both runs. All three hoops are still in her left ear, the headphones are still at her neck, and both gloved hands are still on the decks. Her face is recognizably the same in both, and the room behind her hasn't moved.
The rendering style held too, with the same flat cel shading and line weight across the new tank top and the bare arms that weren't in the original at all. The lightning bolt patch is gone from both, which isn't a surprise, because we replaced the jacket and didn't name the patch, so it left with the garment exactly as it did in the cardigan test.
One change arrived unrequested. In the run with the clause, her eyes are closed. The prompt asked for a grin and said nothing about her eyes, and the model read a wide grin as the kind you make with your eyes shut. It's a reasonable interpretation, but it hides the gold eyes that are part of her design, and the run without the clause kept them open.
Test 4: Pushing an AI Image Editor as Far as It Goes
Every edit so far has left the room alone. Rei's ceiling, wall, and doorway have stayed in place while everything else changed around them. This last test removes that safety net.
We moved her outside, changed the lighting, replaced her jacket, and added text, all in one prompt:
"Move her from this room to an outdoor rooftop party at night with a city skyline behind her, change the lighting to warm string lights overhead, put a black leather jacket on her, and add a neon sign on the wall behind her reading ON AIR"
This was the result we got.
The room became a rooftop, warm string lights replaced the overhead downlights, the bomber became leather, and the neon sign rendered correctly.
Everything landed on the first attempt. The biggest surprise is the sign. Short text is often where image editors slip, but "ON AIR" came back correctly spelled and cleanly placed on the brick wall behind her, looking like a sign that belongs in the scene rather than text floating over it.
The lighting holds together too. The string lights are overhead, and their warm light reaches Rei instead of stopping at the top of the frame. There is a warm edge along her hair and shoulders, warm light across the decks, and the city behind her stays cooler.
Most of the checklist survives. The streak stays on her right, the three hoops are still three, the headphones remain at her neck, and both gloved hands stay on the decks.
Her face looks different at first glance, and the reason is the light rather than a redraw. In the original, she is lit by two cool overhead downlights, and her skin reads as a pale, slightly pinkish white.
On the rooftop, she is lit by warm string lights, and the same skin comes back warmer and a shade deeper. That is what a warm light source should do to a face, and it is the sort of change you want an editor to make rather than drift.
That is the behavior this test exposes. Once the model has to rebuild the room, the light, and the garment at once, everything in the frame gets re-rendered under the new conditions rather than carried across unchanged.
The text should have been the risky part. Instead, it was one of the cleanest, so we pushed further. This time we kept the rooftop result rather than going back to the original, since the neon sign only exists in that image, and asked for two text changes at once while removing the string lights we had just added.
Here's the prompt we used:
"Change the neon sign to read LATE SHIFT RADIO, add a small handwritten note taped below it reading BACK AT 2 AM, and turn the string lights off so only the neon lights the scene"
And this is the result we got.
The neon sign rewrote correctly, and the handwritten note came back legible, while the string lights were removed rather than switched off.
The neon sign handled the change well. "LATE SHIFT RADIO" came back correctly across three lines, still on brick and still glowing. The handwritten note landed too, which was the part we expected to break.
The lighting instruction is where something more interesting happened. We asked for the string lights to be turned off, and the model removed them entirely. There are no wires overhead and no unlit bulbs, just clean sky. The rooftop then re-lit itself around the neon and a lantern, and Rei is now lit from the left, which matches where the sign sits.
That is the same habit we saw with the patch and the streak. When asked to change an object's state, the model acted on the object itself. Turning something off and taking it away produces a similar-looking result at a glance, and it chose the more drastic one. If you want a light source to be dark but still present, say that it stays in the frame.
Simple Vs. Complex AI Image Editing
So, what did our tests actually show? When it comes to Tsubaki.3, instruction following was never the problem. A four-word color swap and a four-part prompt both came back complete, and nothing was ever dropped for being buried in a long instruction.
One instruction was carried out differently than we meant. We asked for the string lights to be turned off, and the model removed them from the scene entirely. It was done, just not as intended.
Preservation was the harder half. The more of the image the model had to rebuild, the more room there was for small changes.
Simple and complex attempts to edit an image with AI, compared side by side across instruction following, preservation, drift, and unexpected changes.
The three-change edit kept the composition intact because every change happened on Rei herself and the room was never touched. The rooftop edit rebuilt the room, the light, and the jacket at once, and that is where her expression softened without being asked.
The takeaway is about how specific a detail is. Smaller details, and anything sitting on an object being replaced, were the ones that drifted. The lightning patch shows this best. It disappeared three times, and every time the jacket underneath it had been replaced. When the jacket stayed, the patch stayed.
Naming it helped, but only when we named it precisely. The streak did the same thing, holding at one when we named it that way and duplicating down the ponytail when we didn't.
So a preservation clause isn't a general instruction to leave things alone. It works when it names the details.
Where Tsubaki.3 AI Image Editing Works Best
After roughly a dozen edits on the same image, we have a good idea of where Tsubaki.3 is reliable and where it needs a careful prompt. Here's an overview of what worked best:
- Fixing one incorrect detail: This is where Tsubaki.3 is strongest. Changing the eye colour affected the irises and left the eyeliner and lashes alone. If an image is already right apart from one detail, that is the easiest edit to make.
- Removing or replacing objects: Taking out the headphones was clean, and the model rebuilt the collar underneath with ribbing that matched the rest of the jacket. The crucial part is that it didn't simply leave an empty space. It filled in what had been hidden in a way that fit the surrounding image.
- Changing outfits and accessories: Both the cardigan and the leather jacket came in without disturbing her face, pose, or hands. The catch is the one we saw earlier. Details attached to the old garment can disappear, so if a patch, pin, or other accessory matters, name it and say where it belongs.
- Making several coordinated changes at once: Three changes and then four both landed in a single prompt. The longer instruction wasn't the problem. Splitting those changes across separate edits would have meant rebuilding the image more than once and could have introduced more drift.
- Changing the environment and lighting together: Moving Rei to a rooftop under warm string lights worked better than expected. The new light actually reached her, with a warm edge along her hair and shoulders while the city stayed cooler behind. The lighting belonged to the new scene rather than looking like an effect placed over it.
The weaker areas were narrower than we expected, and both come down to wording. A trait described loosely leaves room for error. So does an instruction that doesn't say how far to go, which is why turning the string lights off took them out of the scene entirely.
Also, each edit gives you a new render rather than a small patch to the original file, so going back to the original image each time keeps any losses from stacking. If you are new to the platform, the PixAI editing guide covers the basic workflow, and the same idea comes up in character consistency work, where returning to the original reference is more reliable than chaining generations.
So, Can AI Image Editing Change Just One Thing?
Mostly, yes, and the exceptions are predictable. Simple edits were clean. Blue eyes changed the irises and nothing else. Longer prompts weren't the problem either, with three changes landing together and then four, including a neon sign spelled correctly on the first attempt.
What held was anything specific and high-contrast, like her face, the platinum streak, and the three hoops. What drifted was anything described loosely or sitting on an object being replaced.
Reliability tracks how much of the frame has to be rebuilt, not how long the prompt is. So the fix isn't more words. It's naming which details matter, counting them, and placing them.
Try it on something of your own. Upload an image you already like, ask for one change, and check what else moved. Open Tsubaki.3 in PixAI and see what your own edits preserve.
Top comments (0)