I Ran 8 Edits Through an AI Anime Image Editor. Here Is Every Prompt and Result.

Most AI art writing is about generation. This is about the reverse move: taking an image you already have and changing one thing without wrecking the rest. That is what it means to edit anime images with AI, and it is harder than it sounds, because the model has to decide what to touch and, more importantly, what to leave alone.

To see how far it has come, I ran eight edits through PixAI's newly released Tsubaki.3 model across eight image types: a character on a street, an empty room, an event poster, a three-panel manga, black-and-white line art, a photographic-style portrait, a six-panel comic, and a classic shounen action scene. Some were easy. Some I made hard on purpose. For every test you get the exact prompts, base and edit, so you can reproduce them.

Why edit instead of regenerate

When an image is almost right, most people generate again and hope the one wrong detail fixes itself. The problem is that every fresh generation redraws the whole picture.

You finally get the jacket color right, and now the face is different, the pose changed, and the mood you liked is gone.

Editing runs the other way. You hand the model the picture you already have and tell it to change only the wrong part and leave everything else alone.

That makes it the better call when most of the image is already good, when you want a small variation of the same scene, or when you need to fix an object or some text without rebuilding everything.

So the real test for any AI anime image editor is not only whether it makes the change you asked for. It is whether it leaves everything else untouched.

How editing works in Tsubaki.3

You start from an image instead of a blank prompt. Generate or open the picture, click the edit icon, pick the model, describe the change in plain language, and generate.

There is also a dedicated editor if you prefer a full canvas.

The one habit that carried every test was naming two things in each instruction: what should change, and what should stay. The "keep this the same" half is not filler. It is the fence that stops the model from redrawing a face when all you asked for was a new jacket.

Test 1: Relight a scene, keep the character

The change most people want first: shift the whole mood while the character stays herself. This is the hardest kind of preservation, because almost everything in the frame moves.

BASE
an anime girl with shoulder-length teal hair in a high ponytail, round
glasses, wearing a mustard-yellow cardigan over a white blouse, making a
peace sign with her right hand, holding a paper coffee cup in her left hand,
standing on a sunny city street, a small cafe chalkboard sign beside her,
bright afternoon light, anime illustration

EDIT
Change the scene to a rainy evening with wet streets and warm streetlights,
keep the character, her outfit, pose, glasses, and expression the same.

The cafe girl at midday (base left) and after the rainy-evening edit (edit right)

This is about as well as an anime background changer can do. Bright afternoon turned into a wet, glowing night with reflective streets and rain, and she came through unchanged: the teal ponytail, the glasses, the mustard cardigan, the peace sign, even the coffee cup. Framing shifted a little and a few background details moved, but her identity held completely under a total lighting change.

Test 2: Add an object to a detailed scene

Editing is not only for characters. Often the change is an object, and dropping one in beats rebuilding a whole scene. The test is whether the rest of the room survives.

BASE
a cozy cafe interior in late afternoon, empty, worn wooden tables and
mismatched chairs, a glass pastry case catching warm window light, chalk
menu board on brick wall, hanging pendant lamps, potted plants on floating
shelves, a coffee machine with steam, soft golden light from a large
street-facing window, no people, detailed anime background illustration

EDIT
Add a sleeping orange tabby cat curled up on the windowsill, lit by the same
warm window light, keep the tables, pastry case, menu board, lamps, plants,
and all other details exactly the same.

The empty cafe (base left) and after adding the cat (edit right)

I built the base with plenty of surfaces to disturb: wood, brick, glass, metal. The edit handled it. The cat landed on the sill, curled and catching the warm light so it read as part of the scene rather than pasted on, and the tables, pastry case, brickwork, menu board, lamps, and plants stayed put. This is where an AI image editor for anime is at its most useful: one new object added, everything else left standing. The same approach covers when you need to replace an object in an image AI-side rather than add one.

Test 3: Replace text on a poster

Text is where editing gets harder, and I want to be straight about it. AI models still struggle with small or dense text, so a text edit doubles as a limitations test.

BASE
an anime-style summer festival key visual, a girl in a red yukata seen from
behind watching fireworks over a river, festival stalls with lanterns below,
large stylized title "SUMMER FEST" across the top, a subtitle "August 15 -
17" beneath it, a small ticket-price line and a logo badge in the bottom
corner, clean graphic poster layout, anime illustration

EDIT
Replace the main title "SUMMER FEST" with "NIGHT MARKET FESTIVAL". Keep the
new title clearly readable and properly fitted within the original title
area. Preserve the subtitle "August 15 - 17", ticket-price line, logo badge,
all three characters, their clothing, poses and positions, the fireworks,
river, festival stalls, lanterns, reflections, colors, lighting, and overall
poster composition exactly as they are. Do not move, redraw, remove, or alter
any other element.

The poster before (base left) and after (edit right)

The title swap came out well. The longer replacement fit the original title area without overlapping the subtitle or getting cut off, and it stayed readable, which is not a given when a title gets longer. The subtitle, the girl, the fireworks, and the layout all held.

My prompt happens to name three characters, but this poster has one, so she was the only figure to keep. The miss is the small print: the tiny ticket-price line was garbled in the original and stayed garbled after. To edit text in an AI image works for headlines and short lines, but do not expect it to fix tiny type.

Test 4: Edit a single manga panel

A manga page is one image made of several panels, so changing one panel without touching the others is a real preservation test.

BASE
a black and white manga page with three panels and clean gutters, top panel
a wide establishing shot of a rainy school rooftop, middle panel a medium
shot of a short-haired girl gripping the wet railing and looking down, bottom
panel a close-up of her face with a single tear, screentone shading, speed
lines in the middle panel, one empty speech bubble in the top panel, manga
illustration

EDIT
In the bottom close-up panel only, change her expression from a tear to a
small confident smile. Also add the text "I'll tell him tomorrow." inside the
empty speech bubble in the top panel. Keep the other panels, the page layout,
the screentone, the speed lines, and the gutters exactly the same.

The manga page before (base left) and after (edit right)
Both changes landed in the right spots. The tear became a natural, confident smile, and the empty bubble filled with clean, readable dialogue. The three-panel layout, the gutters, the rooftop, the middle panel, the rain, the screentone, and the speed lines all held, and the middle panel was correctly left alone.

The one honest note: the edited panel's face rendered slightly differently beyond the mouth, so it is the same character and framing, just not pixel-identical. The short bubble text came out clean here, the opposite of the tiny poster print, because larger, simpler text is easier to place.

Test 5: Color line art and add a background

This moves an image from one production stage to the next, bare line art to a colored scene. The test is line fidelity: does it color inside the drawing, or quietly redraw it.

BASE
clean black and white anime line art on a white background, no color, no
shading, a girl with long braided hair wearing a hooded traveler's cloak and
holding a lantern, detailed confident linework, full body, character concept
line art

EDIT
Add flat anime color to this exact line art: teal hair, a deep red cloak,
brown boots, and a brass lantern with a warm golden glow. Keep the
character's linework, pose, proportions, facial features, clothing design,
and composition completely unchanged. Do not redraw, reshape, or restyle the
character. Expand the scene by adding a detailed fantasy background around the
character: a narrow forest path beneath tall trees, moss-covered rocks, small
bushes and ferns along the path, a few distant mountain silhouettes, and
subtle atmospheric mist. Add soft moonlight filtering through the trees and
let the lantern cast a warm glow onto the nearby ground, rocks, and lower part
of the cloak. Include a small wooden signpost and a few scattered leaves near
the path. Keep the new background visually secondary to the character. Do not
overlap, cover, or modify the existing character line art. Preserve the
original full-body framing and proportions, and use clean flat anime colors
throughout with simple controlled lighting and no painterly rendering.

The line art (base left) and the colored version (edit right)

This is where the limits showed. The hard part, the background, came out better than I expected: a misty forest path with rocks, ferns, a signpost, and a warm lantern glow. But it went past what I asked for on style. Instead of the flat colors I specified, it added shading, atmospheric light, and a painterly finish, and the cloak and proportions drifted a little from the original lines. It colored and rendered the piece rather than flat-filling it. If you need to preserve an exact production stage, plan around that.

Test 6: Relight a portrait and add a second character

The most complex edit so far: a full relight, a new person, new objects, and the original subject kept exactly. The base needs flagging first, because it revealed a real limitation.

BASE
a photorealistic portrait of a young woman standing by a large rain-streaked
window, soft overcast daylight, shallow depth of field, fine film grain,
realistic skin and hair texture, muted natural color grade, photographic

I asked for a photorealistic portrait, and this is the one place the model plainly missed. Despite the prompt, Tsubaki.3 returned an anime-style illustration with stylized features rather than real skin and film grain. The rainy window and soft light were right, but the core photographic requirement was not met. Useful to know going in: asking this model for a true photo is not its strength.

EDIT
Relight this image to a warm golden-hour sunset streaming through the
rain-streaked window, with soft orange sunlight, natural highlights, gentle
shadows, and a subtle lens flare. Increase the shallow depth of field so the
background falls softly out of focus. Add a second young woman standing near
the window in the background, slightly behind and to the right of the original
woman. Give her shoulder-length brown hair and a simple cream-colored sweater,
with a natural relaxed expression as she looks toward the window. Expand the
environment with a small wooden table beside the window, a ceramic coffee cup
on the table, a few books, a potted green plant, and a warm table lamp. Keep
the rain visible on the window and add subtle reflections of the sunset and
interior objects in the glass. Most importantly, keep the original woman
exactly the same. Preserve her face, eyes, hair, pose, hands, clothing,
proportions, and position. Do not redraw, reshape, replace, or restyle her. Do
not alter her expression or identity. The newly added character and objects
must fit naturally into the existing scene without covering or modifying the
original subject.

The portrait before (base left) and after the relight and added character (edit right)

The edit itself was strong, which makes for an odd but honest split. The golden-hour relight worked: warm sunlight through the window, a visible lens flare, stronger depth separation, and convincing reflections on the wet glass. It added the second woman and the table, cup, books, plant, and lamp without crowding the frame, and blurred the new character correctly to match the shallow depth of field.

The original woman stayed recognizable, her face, hair, pose, and position held. The caveat: the strong warm light re-rendered her skin, hair, and clothing, and her fine facial details came out softer than in the original. A big, multi-part edit largely worked, even though the base it started from had missed its own brief.

Test 7: Change one detail across a six-panel comic

The most demanding preservation test. I generated a full six-panel story, then asked to change one small character detail across the entire page, which means the edit has to hold in six drawings of the same person at once.

BASE
Create a polished Japanese anime manga page consisting of 6 sequential panels
telling one coherent short story.
Main character: a 17-year-old Japanese high school girl with short dark-blue
bobbed hair, straight bangs, amber eyes, a small beauty mark beneath her left
eye, and a distinctive red ribbon tied around her school uniform collar. Keep
her facial features, hairstyle, body proportions, uniform design, and red
ribbon exactly consistent across every panel.
Story: On her way home after school, she notices a small abandoned kitten
hiding beneath a vending machine during a sudden evening rain. She initially
hesitates, then kneels down, reaches toward the kitten, shields it from the
rain with her school bag, and finally smiles as the kitten approaches her.
Panel 1: Wide establishing shot of a Japanese residential street after school,
dark clouds gathering, the girl walking alone with her school bag.
Panel 2: Medium shot as she notices the kitten beneath a vending machine,
surprised expression.
Panel 3: Close-up of her face looking concerned, followed by a detailed view
of the frightened kitten.
Panel 4: Dramatic low-angle shot as heavy rain begins, with the girl kneeling
and extending one hand toward the kitten.
Panel 5: Dynamic composition showing her using her school bag to shelter the
kitten from the rain, with strong diagonal rain lines and dramatic lighting.
Panel 6: Warm emotional close-up as the kitten finally approaches her hand and
she smiles.
Use varied camera angles and shot sizes across the page. Use high-quality
Japanese manga aesthetics: clean expressive anime linework, detailed character
rendering, screentone shading, speed lines where appropriate, cinematic
lighting, strong panel composition. Add natural Japanese dialogue and sound
effects in the speech bubbles, legible and correctly integrated, short and
natural rather than excessive.

EDIT
Change the character's red ribbon to blue in all six panels. Keep her face,
hair, uniform, the kitten, the panel layout, the rain, the backgrounds, the
Japanese text, and the manga style exactly the same.

The six-panel page before (base left) and after the ribbon edit (edit right)

This was the standout. The ribbon turned blue in every one of the six panels, and her identity, the panel layout, the rain, the backgrounds, and the manga style all held. Changing one character detail consistently across six separate drawings is a hard consistency test, and it passed with only very minor differences in facial linework between versions.

The weakness was in the original generation rather than the edit: some Japanese dialogue and sound effects read as unnatural, the same small-text problem from the poster.

Test 8: Rebuild a fight scene around an unchanged character

The hardest of all: keep a character identical while replacing the entire scene and adding a second character mid-action. I started with a classic shounen warrior on a cliff at sunrise.

BASE
A young anime warrior standing on a cliff edge at sunrise, spiky black hair
with sharp defined strands catching orange light, determined narrow eyes with
sharp highlight points, wearing a torn orange and blue gi with a red sash,
scar across one cheek, fists clenched, wind blowing his clothes and hair to
one side, small energy aura beginning to flicker faintly around his shoulders.
Background: a rocky mountain cliff overlooking a valley shrouded in morning
mist, distant peaks against a fiery orange-to-yellow gradient sky, a few birds,
sparse windswept pine trees. Style: classic shounen anime, bold clean
linework, flat cel-shading with hard-edged shadows, vibrant saturated colors,
dramatic sunrise lighting with strong rim light, sharp facial features
reminiscent of 90s-2000s Toei animation.

EDIT
Keep the same character exactly as generated: same face, same spiky black
hair, same sharp eyes, same scar across the cheek, same torn orange and blue
gi with red sash, but transform the scene entirely. Transplant the character
into a mid-air battle stance against a new villain: a tall armored antagonist
with pale white skin, long silver hair whipping violently, glowing violet
eyes, a cracked black and crimson breastplate with jagged shoulder spikes, one
clawed hand raised and crackling with dark purple energy, a torn cape, cold
sneering expression. Both are locked in a clash mid-motion: the hero's fist
connects with a shockwave of golden energy against the villain's clawed hand,
impact distortion rippling outward, clothes and hair whipped by the collision,
debris frozen mid-air. Environment: replace the sunrise cliff with a
collapsing ancient stone temple ruin at night, shattered pillars cracking
apart in the air, glowing blue runic energy from fractures in the floor, a
massive cracked moon low behind the ruins, storm clouds and lightning, embers
drifting. Style: same classic shounen cel-shaded look, bold linework, flat
shading, high saturation, shifted to a high-contrast night battle: deep blues
and purples, gold and violet energy, dramatic uplighting, motion blur on both
characters' outer limbs.

The shounen hero (base left) and after the fight-scene transformation (edit right)

The base nailed the shounen look: bold linework, flat cel-shading, spiky hair with rim light, the torn gi, though the energy aura came out softer than the flickering effect I asked for. The edit is the impressive part. The hero's face, hair, scar, and gi held intact while the whole scene changed to a collapsing night temple with a fully detailed villain mid-clash, floating debris, and a golden-versus-violet energy impact. Locking identity through a total scene swap is the hardest thing an edit can do, and it held. Only the motion blur was subtler than I wanted, so the impact reads a touch soft.

One honest flag, unrelated to model quality: the hero design lands very close to an instantly recognizable existing character rather than an original one. That is fine for personal testing, but it is a copyright and platform risk if the image goes public, and the edit only doubles down on it. Prompt for a famous style and this is the trade you make.

Instruction editing vs inpainting

Everything above is instruction-based editing: you describe the change in words and the model works out where to apply it.

The older approach is good to know too. Inpainting has you manually mask the exact area you want changed and regenerate only that patch. Neither wins across the board. Instruction editing is faster, reads naturally, and handles whole-scene changes like the Test 1 relight that would be awkward to mask by hand.

Inpainting still wins when you need tight local control over one small region and do not want the model deciding the edges. The inpainting and outpainting guide covers that route, and AI inpainting for anime remains a solid choice for pinpoint fixes.

What I would tell you before you start

None of these lessons are complicated, and they all came out of the eight tests. Name what should change and what should stay in the same instruction, because the "keep this the same" half protects the parts you like. Change one thing at a time, since stacking several edits into one prompt is where results get messy.

Be specific about clothing, objects, gestures, and expressions. If you want to change a pose in an image AI-side, the same rule holds: name the new pose and lock the rest.

Expect small or dense text to stay rough, and lean on edits for headlines, not fine print. And always compare the result against the original, because the change you did not ask for is the easy one to miss.

The takeaway

Editing is the tool for when you already have an anime image you want to keep. Instead of regenerating and gambling the whole picture, you change the one wrong part and leave the rest standing, which keeps the version you worked for and costs far less.

Across these eight tests Tsubaki.3 did that well more often than not: relighting a scene, adding an object, swapping a title, editing one manga panel, coloring line art, relighting a portrait and adding a character, changing a detail across a whole comic, and rebuilding a fight scene, all while holding the surrounding image steady.

It is not perfect. Small text stays rough, a strict style instruction can drift, a heavy relight softens fine detail, and it would not give me a real photo when I asked. But for targeted changes to characters, objects, text, and backgrounds, it does the job editing is supposed to do. If you have an image that is almost right, try editing it on PixAI instead of starting over.