I went into this expecting the pictures to be the challenge. They were not. Across eight scenes, the composition scored above 9 more often than not. The words in the bubbles were where things came apart.
That split is the whole story of using an AI manga generator today, so I want to organize this review around it rather than march through the scenes one by one. A good anime picture and a real manga panel are different problems. You can prompt a beautiful character all day, but drop a speech bubble on top and you still have an illustration with text on it, not a page from a comic. Manga needs a moment, a character framed for that moment, a line that belongs to them, and screentone and speed lines doing the emotional work.
So I spent a week putting Tsubaki.3 on PixAI through eight scenes, from a quiet rooftop betrayal to an all-out tavern brawl, to see how well you can create manga with AI right now. I grouped what I found by the questions I was really asking. Every prompt and score is below.
The pieces that make an image manga
Before the tests, the target. A few things separate a manga panel from a nice anime drawing. It shows a moment, not a pose, so something is happening in the frame. The framing supports that moment, a close-up for emotion or a wide shot for chaos. The speech bubble belongs to a specific character, with its tail pointing at whoever is talking. And the manga tools carry the mood, the screentone, the speed lines, the heavy ink blacks.
Every prompt below aims at those four things at once. That is the part a plain character generator cannot do, and the part that makes this hard.
How I set the tests up
The workflow is the same each time. You describe a scene, not just a character. I lead with what is happening, then define who is in the frame, then set the camera, then add a short line of dialogue with the speaker named, then call in the manga tools. Everything ran on Tsubaki.3, PixAI's newest model. If your prompts keep coming out flat, the PixAI prompt formula is a good starting point.
One note that applies to every scene below: I generated each prompt first on base Tsubaki.3 with default settings, no style and no LoRA, and that is the left image. Then I ran the exact same prompt again with one or more LoRAs added, for the right image. The prompt never changes between the two, only the LoRA does.
Question 1: Can it frame a single dramatic moment?
This is the friendliest case, one character and one line, and it is where the model is strongest.
The first scene was a lone mechanic kneeling beside a wrecked robot in a burning hangar, looking up in disbelief.
a dramatic black and white manga panel, a young mechanic girl with
grease-smeared cheeks and goggles pushed up into her wild hair, kneeling
beside a half-destroyed giant robot in a burning hangar, sparks and smoke
around her, clutching a wrench, looking up in disbelief at something
off-panel, debris frozen mid-air, heavy screentone, sharp speed lines,
dynamic low angle, a trembling speech bubble reading "...You're alive?",
clean manga lettering, high contrast, manga illustration
On the base model, with no help at all, it understood the assignment. Heavy blacks, real screentone, a mecha with weight, and the girl there with her goggles, wrench, and grease. It looks like a panel, not a character card, and the line came through spelled right and legible in the corner. Two misses: the hangar was more smoke than fire, and her face looked surprised rather than devastated. 8.5 out of 10 on the first try.
Left base Tsubaki.3. Right with LoRAs.
The second scene pushed the same idea into pure framing, two very different single-character panels.
a black and white manga action panel, a martial artist mid-leap throwing a
punch straight at the viewer, extreme foreshortening, radial speed lines
bursting from the fist, heavy screentone, dynamic diagonal composition, a
jagged shout bubble reading "HAAA!", manga illustration
a black and white manga close-up, a scarred old swordsman grinning slowly in
the shadows, one eye glinting through his messy hair, cracked screentone
texture, heavy ink blacks, ominous mood, a small sharp speech bubble reading
"Finally. A real one.", manga illustration
The first is a fist coming straight at you, extreme foreshortening, speed lines off the knuckles, a jagged "HAAA!". It lands the way an action page should. The second goes the other way, a scarred swordsman grinning in the shadows, heavy blacks, cracked screentone, one eye catching the light. "Finally. A real one." Quietly menacing. Both scored 9.5, and both lines spelled right. One character, one focus, one short line. This is Tsubaki.3 at its best.
Left base Tsubaki.3. Right with the Manga Style LoRA test 4
Question 2: Can it hold several characters in one frame?
Harder, because now the model has to keep separate identities from bleeding together and tail each bubble to the right speaker.
A two-character rooftop betrayal was the cleanest result of the set.
a black and white manga panel, two students on a rooftop at sunset, on the
left a tall girl holding out a folded note with a calm unreadable face, on
the right a shorter boy frozen with a shocked expression and a single sweat
drop, wind scattering cherry blossoms between them, the city far below, soft
screentone and delicate linework, a small speech bubble from the girl reading
"I already know.", clear manga lettering, balanced two-shot composition,
manga illustration
This came out at 9 out of 10. The calm girl on the left, the frozen boy on the right, distinct at a glance, nothing blurring together. The bubble landed between them and pointed at the girl. "I already know." A whole relationship in three words. The only real miss was the sunset, which came out almost white with sparse blossoms. A pass with the Manga Style LoRA nudged it to 9.2, sharper ink and a more deliberate look.
Left base Tsubaki.3. Right with the Manga Style LoRA.
Then I scaled the cast up. A three-character brawl was the hardest scene I threw at it.
a lively black and white manga panel, three young adventurers caught in the
middle of a chaotic tavern brawl: a determined girl in the center swinging a
wooden chair at an attacker off-panel, a terrified boy behind her ducking under
a flying mug with both hands over his head, and a smug older woman sitting
calmly at the table drinking tea while chaos erupts around her, overturned
chairs, flying mugs, scattered coins, smoke and motion lines, other patrons
scrambling in the background, exaggerated facial expressions, dynamic
overlapping poses, heavy screentone, bold ink blacks, energetic speed lines,
dramatic perspective, three speech bubbles: the girl shouting "Get down!", the
boy yelling "I'M TRYING!", and the older woman calmly saying "Quite noisy.",
clear manga lettering, each bubble clearly tailed to the correct speaker,
manga illustration
It held together. Flying mugs, overturned chairs, speed lines everywhere, and nothing melted into a blur. All three characters stayed distinct, and all three lines came through correctly, which is a real result in a frame this busy. The miss is precision: I asked for the girl to swing a chair at an off-panel attacker, and there is no chair anywhere. It got the feel of a brawl, not the exact choreography. The Manga Style LoRA sharpened the rendering and, more usefully, tightened the bubble-to-speaker matching. 9 on the base, 9.2 with the LoRA.
Left base Tsubaki.3. Right with the Manga Style LoRA test 7
A four-adventurer tavern argument scored the same 9.5 on composition, four distinct characters around a table packed with maps, mugs, coins, and lanterns, and three bubbles sized right with the knight getting the biggest one. Its text is a story for the dedicated section below, because that is where it stumbled.
a lively black and white manga panel, four adventurers arguing around a
torch-lit tavern table covered in maps, mugs, and coins: a big armored knight
slamming his fist on the table and shouting, a sly hooded rogue leaning back
with a smirk and boots on the table, a nervous young mage clutching a staff and
sweating, and a calm elf archer with crossed arms and closed eyes. Wooden
beams, hanging lanterns, and a crowd of blurred patrons in the background,
heavy screentone, detailed linework, dynamic composition. Three speech bubbles:
a large jagged bubble from the knight reading "We attack at dawn!", a small
smug bubble from the rogue reading "Or we don't.", and a tiny nervous bubble
from the mage reading "...guys?". Clear manga lettering, each bubble clearly
tailed to the correct speaker, readable group composition, manga illustration
Left base Tsubaki.3. Right with the Manga Style LoRA test 5
The lesson across these three: as a multi-character AI comic generator, Tsubaki.3 keeps identities separate and staging believable even when the frame is packed. It gives up some exact choreography as the scene gets busier.
Question 3: Can it keep a character across a sequence?
One panel is one thing. A strip that has to hold the same face across several is the real test for consistent manga characters. So I asked for a 4-koma, four panels telling one joke.
a black and white 4-koma manga strip, four stacked panels with the same
deadpan girl throughout: panel 1 she confidently casts a summoning spell with
a glowing circle; panel 2 smoke everywhere, she looks hopeful; panel 3 a tiny
unimpressed cat sits in the circle staring at her; panel 4 she deadpans at the
reader while the cat licks its paw. Consistent character across all four
panels, screentone, clean panel borders, short dialogue and sound effects in
bubbles, manga illustration
The gag lands. She casts a spell, smoke fills the panel, a tiny unimpressed cat appears in the circle, and she deadpans at you. Four beats, one joke, no explanation needed. The girl stayed the same across all four panels, and so did the cat. That is the part that matters here, and it held. The Manga Style LoRA lifted the look to 9.5. The dialogue, though, is the sticking point, and it belongs in the next section.
Left: base Tsubaki.3. Right: with the Manga Style LoRA.
Question 4: Can it work in full color?
Manga is not always monochrome. Webtoons run in full color, so I gave it two color scenes.
The first was a soft vertical webtoon panel.
a full-color webtoon style vertical panel, soft digital coloring with clean
lineart and gentle cell shading, a girl with pastel lavender hair and freckles
sitting alone on the steps of a cozy cafe at golden hour, holding a warm drink
in both hands, string lights and potted plants around her, a sleepy orange cat
curled beside her, warm bokeh light, soft emotional atmosphere, gentle color
grading, a rounded speech bubble above her reading "Maybe today's okay.", clean
webtoon lettering, vertical phone-friendly composition, Korean webtoon aesthetic
It came out at 9.5. A lavender-haired girl outside a cozy cafe, sleepy cat, golden light, warm and clean and shaped for a phone. Exactly the webtoon look, and "Maybe today's okay." spelled right. A pass with the Manga Style and Clear Style LoRAs got cleaner and sharper, though the girl came in closer to the camera and lost some of the quiet, alone framing of the base.
Left base Tsubaki.3. Right with the Manga Style and Clear Style LoRAs.
The second was a full-color splash page, the big dramatic kind.
a dramatic full-color manga splash panel, a young swordsman standing alone on a
ruined bridge during a violent thunderstorm, his torn coat and hair whipping in
the wind, katana lowered at his side, rain pouring across the scene, lightning
illuminating a massive shadowy creature emerging from the mist behind him,
broken buildings and debris scattered across the river below, the swordsman
looking back over his shoulder with a determined expression, deep atmospheric
perspective, foreground rain streaks, dramatic lightning, rich but controlled
manga coloring, detailed linework, selective cel shading, strong contrast,
cinematic composition, a small speech bubble near him reading "So you finally
came.", clean manga lettering, full-color manga illustration
The base pass came in at 9.3. A swordsman on a ruined bridge in a storm, lightning behind him, a huge creature in the mist. Controlled color instead of blown-out anime, real depth, and "So you finally came." came through clean. The LoRA pass is a cautionary tale I will get to under LoRAs, because it is the clearest example of a good LoRA wrecking a brief.
Left base Tsubaki.3. Right with the Manga Style and Body Aesthetics LoRAs.
The takeaway: color is not an afterthought here. Both color scenes scored above 9, so this doubles as a full-color AI comic generator, not a black-and-white-only tool.
The real limit: AI manga text
Here is the pattern that ran under everything. The pictures were strong. The words were the weak point, and the failure is consistent enough to plan around.
Two scenes show it plainly. On the 4-koma, the joke and the character held, but the dialogue came out in Japanese, convincing squiggles in the wrong language, because I had not locked English. On the tavern argument, the knight's line was supposed to say "We attack at dawn!" What came out was "We atteck ot dawn!" Attack became atteck, and at became ot. I ran it three times, including with the Manga Style LoRA, and got the same error every time.
So for the AI speech bubble generator side of this, the rules are simple. Always specify English. Keep your lines short. Keep the words common. Long or unusual words trip it up, and you should plan to redo the text on another pass. Short, plain lines like "I already know." and "Finally. A real one." spelled right every time. That is where AI manga text stands right now, close on common phrases, unreliable on dense or rare words.
The LoRA that helped, and the ones that hurt
The other cross-cutting lesson: the right LoRA supports the brief, and the wrong one quietly overrides it.
The Manga Style LoRA was the good one. It sharpened every black-and-white scene, heavier ink, denser screentone, more deliberate lines, and it backed up the manga look instead of fighting it. Keep that one for manga.
The character LoRAs fought the prompt. On the mechanic panel, stacking Crystal Eyes, Body Aesthetics, and Anime Girl strengthened the ink but colored the eyes inside a black-and-white manga and turned my grease-covered mechanic into a glossy heroine.
On the color splash page, running Body Aesthetics alongside Manga Style handed me a woman when the prompt clearly asked for a young swordsman, a great-looking panel that missed the most basic word in the brief.
Clear Style was the exception, cleaner and sharper on the color webtoon and otherwise out of the way.
So match the LoRA to your goal. A manga LoRA for manga. Skip the eye and body LoRAs unless you want a pin-up rather than the character you described.
Keeping a character consistent
If you are building a series instead of one image, your character has to stay recognizable from panel to panel. The 4-koma showed the model can hold a face and hair four panels deep on its own.
For a recurring character across a longer story, feed it a reference image to anchor to. PixAI's Reference Pro guide covers that workflow, and the more panels you make, the more it counts.
Field notes
A few habits came out of the eight scenes. Describe the moment, not the outfit, and lead with the action. Keep dialogue short, since one line beats a paragraph. Name the speaker right in the prompt. Always specify English for the text. Give a camera angle when it matters, close-up, wide, or low.
Pick one or two manga tools instead of stacking all of them. Simplify when the frame gets crowded, because fewer elements come across cleaner. And expect to redo longer text on a second pass.
The verdict
So, does it make manga? Yes, and more convincingly than I expected. It thinks in panels, frames a moment well, holds a busy scene without coming apart, keeps a character consistent across a strip, and handles full color as easily as black and white.
That is a real AI manga panel generator, not an anime app with speech bubbles bolted on.
The limits are just as real. Long dialogue garbles, rare words come out wrong every time, and exact choreography goes loose in busy panels. The wrong LoRA will hand you a lovely image that ignores what you asked for. None of that stops you.
You plan around it, with short lines, the right LoRA, and simpler scenes when precision counts.
Making manga with AI was never about generating one pretty picture. It is about pulling a moment, a frame, a character, and a line together in a single shot, and Tsubaki.3 does that, flaws and all.
So instead of stitching panels together by hand, you can build the whole thing in one generation. Open Tsubaki.3 on PixAI and make a single panel of your own. You will learn more from that one attempt than from anything I could write here.
Top comments (0)