Editing an existing image with AI beats starting over — provided the model does what you asked and nothing more. Making the requested change is only half the job. The other half is leaving untouched everything you never mentioned.
New hairstyles alter faces. Background swaps bleed into lighting and composition. Added objects shift position between generations. So this AI image editing test set out to find how well Tsubaki.3 copes with progressively harder edits while holding onto the original image.
Working from a single seated anime character as the source, I ran the Tsubaki.3 model on PixAI through a simple object replacement, a structural character edit, a three-part instruction, and finally a large environmental change.
Whether Tsubaki.3 could edit image with AI at all was never the question.
The real one: can it make the requested change and leave everything else alone?
PixAI interface set up for image-to-image editing using the Tsubaki.3 model.
What Are We Testing?
Any useful AI image editor has to do two jobs simultaneously:
- Instruction following: Did Tsubaki.3 complete everything the prompt asked for?
- Preservation: Did the character, pose, composition, lighting, style and unrelated details remain reasonably consistent?
For each prompt I reviewed several generations, picking the primary result for the before-and-after comparison plus a key supporting output to show compositional behaviour. Every edit got judged on how accurately it delivered the requested changes while leaving unrelated elements alone.
To keep things consistent, each test started from the same original seated male character in a bamboo room rather than editing the previous result. That way each instruction was applied to identical source material.
Test 1 — Simple Targeted Editing
Deliberately easy to begin with.
I wanted to see how Tsubaki.3 handled one clearly defined object change while preserving character and environment.
Before vs. After
Left: Source. Right: Edited Image.
Prompt
"Change the wooden practice sword to a red lacquered folding fan held in his hand, keeping his pose, pale green kimono, facial expression, and bamboo background unchanged."
Evaluation
Tsubaki.3 swapped the wooden practice sword for a red folding fan and left the character, pose and bamboo background largely intact.
A strong baseline for targeted AI image editing, then. The requested object changed without disrupting the rest of the scene.
Verdict: Strong instruction following; good preservation.
Test 2 — Structural Editing
Next, something beyond object replacement — two changes to the character himself.
Before vs. After
Left: Source. Right: Edited Image.
Prompt
"Change his ponytail to loose, shoulder-length hair and give him a stern expression, preserving his kimono, face structure, skin tone, and tea house porch."
Evaluation
Both instructions landed: the ponytail became loose shoulder-length hair, and the calm expression turned noticeably stern. General face structure, kimono and setting all stayed recognisable.
Structural changes clearly demand more reconstruction than swapping an object. The model can still deliver them while preserving the broader character — a useful AI image editor preserve details test.
Verdict: Good instruction following; character identity and scene preserved reasonably well.
Test 3 — Complex Editing: Three Changes at Once
The core AI image editing test.
Rather than an isolated change, one prompt combining three distinct edits: clothing, environment, atmosphere.
Before vs. After
Left: Source. Right: Edited Image.
Prompt
"Change the pale green kimono to a dark crimson haori, change the wooden porch to stone steps, and remove the morning sunlight to make it overcast, keeping his face and hair intact."
Evaluation
One of the stronger showings for instruction following.
All three changes came through clearly. Pale green kimono became a dark crimson layered haori, stone steps replaced the wooden porch, and the bright morning atmosphere shifted toward something darker and overcast.
Face and hair held up well — important, since the prompt explicitly named them for preservation.
The notable change was the whole image shifting in tone to accommodate the new environment and clothing. The result doesn't read as three isolated edits pasted onto the original; Tsubaki.3 treated them as one scene.
Which is a strength when you want a coherent final image.
It's also where preservation gets complicated.
The three instructions were followed, but new clothing and environment inevitably moved the overall visual balance. The scene turned darker and more dramatic, and some details around the character got interpreted differently.
Targeted AI image editing working at a higher level, then — just not perfectly isolated.
Verdict: Strong multi-instruction following; good character preservation, with expected visual drift caused by the larger scene changes.
Test 4 — Push the Edit Further
For the last stage, a much larger edit.
Rather than changing the character or adding a small object, I asked Tsubaki.3 to transform the surrounding environment while explicitly preserving face, pose and framing.
This is where AI image editor preserve details becomes genuinely hard.
Before vs. After
Left: Source. Right: Edited Image.
Prompt
"Change the background to a snowy Zen temple courtyard with falling snowflakes, preserving the exact medium-shot camera distance, sitting pose, facial structure, and hand placement without zooming out into a wide full-body composition."
Deliberately specific, that instruction.
I wasn't only requesting a new environment. I was spelling out what should not happen to the composition.
Evaluation
The result tracked the prompt closely. Bamboo gave way entirely to a detailed Zen temple courtyard — snow-covered buildings, trees, falling snow — while medium-shot camera distance, sitting pose, facial structure and hand placement stayed consistent with the source. Even the explicit no-zoom-out instruction held.
That combination is the best possible outcome for a full environmental overhaul: everything around the character changed, everything about the character and composition stayed put.
Alternative generation from the same batch: camera pulled back into a wide, full-body shot.
Worth flagging: one of the four generations in the batch broke that line, pulling back into a wide full-body shot rather than the requested medium shot. A useful reminder for AI image editing workflows — on full environment rewrites, check each generation individually instead of assuming the first result matches your framing, though this one wasn't representative of the batch.
Verdict: Excellent environment transformation, with framing, pose and facial structure all preserved in the selected result; occasional framing drift elsewhere in the batch is worth checking for.
Simple vs. Complex Editing — What Changes as the Prompt Gets Harder?
A clear pattern emerged across the four: difficulty tracked how interconnected the requested changes were, not how long the prompt was. Which says something useful about AI image editing — the more elements a change touches, the harder preservation gets.
- Local Edits (Sword to Fan): Highly controlled with minimal disruption to the rest of the scene.
- Structural Edits (Hair & Expression): Alters face/head geometry, introducing minor localized variation.
- Coordinated Edits (Kimono + Porch + Atmosphere): Successfully completed, but naturally forces lighting and tone across the whole image to adjust.
- Large Scene Transformations (Snowy Courtyard): Successfully completed, with tight framing held in the selected result, though batch consistency is worth double-checking on full rewrites.
What was most likely to drift?
The most variable elements across these tests:
- Object placement
- Hair/expression details
- Lighting and colour
- Camera framing during major environment changes
Broad character identity proved far more stable than any of these smaller details.
Useful to know if you're working out whether an AI image editor preserve details reliably enough for your workflow.
Explicit preservation instructions plainly helped define the task without guaranteeing every detail survived identical.
For especially sensitive edits I'd therefore split a complicated request into several steps. Practically, that means comparing result against source side by side rather than trusting a glance — particularly around hands, hairlines and camera framing.
Tsubaki.3 AI Image Editing Test Results at a Glance
Where Does Tsubaki.3 Editing Work Best?
Having run these, I'd sort Tsubaki.3's strongest AI image editing use cases into a few categories.
Fixing or replacing individual details
The folding-fan test showed Tsubaki.3 replacing an object while keeping the broader image stable.
Handy for fixing an element that isn't working without regenerating everything.
Controlled character variations
Hair and expression demonstrated that meaningful character changes are possible while basic identity survives.
Useful for building different versions of one character across an illustration, comic or visual concept.
Multiple coordinated edits
The three-part edit was among the clearest demonstrations that several related instructions can go into one prompt.
Valuable when changes belong together — clothing, environment and lighting shifting to create a different version of the same scene.
Full environment rewrites
The snowy Zen temple test showed Tsubaki.3 rebuilding an entire scene — background, weather, lighting — while holding the character's pose, facial structure and camera framing firmly. That's what makes large environmental changes usable at all: everything around the character can change without character or composition drifting.
One thing to watch: with so much of the image being regenerated, generate a few options and check framing against the original rather than assuming every generation in a batch holds a constraint as tightly as the selected result did.
Final Verdict
So — can an AI image editor change one thing without changing the rest?
Across these tests Tsubaki.3 handled local and multi-instruction edits remarkably well, and held up through a full environmental rewrite in the selected result.
Which leads to the main takeaway from this AI image editing test: Tsubaki.3 is good at understanding what you want changed. It's less consistent about understanding exactly where the boundary of that change should stop.
For targeted AI image editing, the best results came from defining both sides clearly — what to change and what to preserve.
For AI anime image editing specifically, that makes Tsubaki.3 a useful option for fixing individual details, creating character variations, adding objects, changing environments, and making several coordinated edits without starting over.
Just don't read "preserve everything else" as a promise that everything else stays pixel-for-pixel identical. The more interconnected the edit, the more closely you should compare result against original, and the more useful multiple generations or smaller editing steps become.
That's ultimately the difference between an AI image editor that can make an edit and one that can make a controlled edit.
Want to test this yourself? Start with a small object change — one clear instruction. Then push further, adding constraints and complexity a step at a time. The real test isn't whether the edit works. It's what happens to everything you didn't mention. Try Tsubaki.3 on PixAI and see what holds up.
Top comments (0)