Tired of your OC changing every generation? See how character reference AI keeps them recognizable across new scenes, outfits, and poses in PixAI.
In a traditional anime studio, an animator doesn't usually draw a character just from memory. They work from a model sheet, a locked reference showing the design from several angles with every detail that has to stay the same. Without it, a character slowly turns into someone else across dozens of episodes.
AI image generators have the same limitation, and it is the gap character reference AI is built to close. No consistent anime character generator can carry a design from one image to the next on its own. Every time you hit "Generate", the model builds your character again from whatever your prompt says.
We ran into this while designing an OC named Ren, a young wizard with dark navy hair streaked pale over one eye, round gold glasses, and heterochromia, amber on one side and dark on the other.
Ren, our original character, standing in a library with his owl on his left shoulder.
The first image came out exactly right. Then we wanted a second one, and that is where things got complicated. Same wizard, different room, same prompt, and the boy who came back had his eye colors on the wrong sides, a cloak that had redrawn itself, and an owl that was suddenly a different bird.
An OC rarely appears once. Comic panels, story illustrations, seasonal art, and VTuber promo assets all need the same character to hold across dozens of images, and a prompt can't do that on its own.
There is another way, closer to what the studio does. Character reference AI works from the image you already have instead of rebuilding him from text, and PixAI, an anime-focused AI art generator, has it built in.
In this article, we'll see how far one reference image can go in PixAI, generating the same character in different scenes, new outfits, and formats his original prompt never described. Let's get started!
Why Is It Hard to Generate the Same Anime Character Consistently?
A prompt is a description, and descriptions leave room. "Navy hair with a pale streak, one amber eye and one dark, green cloak, owl on his shoulder" is a fair account of Ren, and it is also satisfied by thousands of different boys. All of them are technically correct. Only one of them is him.
That gap is where character drift lives. An AI model isn't remembering your character and drawing him again. It reads your words and builds someone new who matches them, and the parts you never put into words are the parts it invents.
Broad traits usually survive. Hair color, eye color, the general shape of an outfit. What slips is everything underneath, like the way the fringe splits, the set of the eyes, the width of the jaw. You would spot those instantly in a finished image, but you never wrote them down.
Counted details go next. Three buttons become two, embroidered cuffs simplify into plain ones, and an owl comes back as a different bird.
Difficulty also scales with what you change. A new background is easy, because the character is left alone. A new pose is harder. A new outfit or camera angle is hardest, since the model has to rebuild the parts you were relying on to recognize him.
Our OC Ren on the left next to three prompt-only generations, where the eye colors swap sides, the cloak redraws itself, and the owl comes back as a different bird each time.
None of this shows up in a single illustration. It shows up the moment a character has to appear more than once, which is the entire point of an OC.
Prompt-Only Generation vs. a Character Reference AI
The three images above are what prompt-only generation gives you. Not failures exactly, since every one followed the instructions. The instructions were just never complete.
A prompt carries what you wrote down. What it can't carry is which side the amber eye sits on, how wide the jaw is, or what kind of owl you had in mind. Those live in the image you already made, and nowhere in your text.
That is the crucial difference. A prompt describes what your character should look like. An AI character reference generator works from what your character already looks like.
A comparison of prompt-only generation and character reference AI across what each one works from, what it holds, and where it tends to drift.
Before you have a design you like, prompting is how you find the character. After that, it is how you can lose him.
However, a reference doesn't guarantee consistency. It gets you closer than text does, and next we'll walk through the tests that show where that holds and where it slips.
Setting Up a Character Reference AI Workflow with Tsubaki.3
We chose Tsubaki.3 to run these tests and generate the same character in different scenes. It is PixAI's newest image model, built with a stronger focus on control and consistency, and character reference is one of the things it handles directly. It accepts up to three reference images, either pulled from your own PixAI generations or uploaded from your files.
The setup is short. Select Tsubaki.3, add Ren's image as the character reference, and write a prompt that covers only what should change. In this case, we asked for one new detail, a lit candle in his hand.
The Tsubaki.3 model selected in the Generate tab with Ren loaded as the character reference and a seven-word prompt in the field.
The prompt was simple. Everything the model knows about Ren came from the image, so the only thing to check is whether he came back as himself with a candle added.
Ren on the left and the seven-word candle prompt on the right, with the eye colors, streak, buttons, embroidery, and owl all carried over.
That is the rule for everything that follows. Each result gets checked against the reference, never judged on its own. An image can be well drawn, well lit, and still be the wrong person, and that is exactly the failure this workflow is meant to catch.
Here is what we checked on Ren every time:
- Amber eye on the left, dark eye on the right
- Pale streak in the hair and which side it falls on
- Round gold wire-frame glasses
- Three bronze buttons
- Pale scrollwork embroidery on the lapel and cuffs
- The owl, and whether it comes back as the same bird
Six things, and four of them are countable or have a side, which is deliberate. Vague traits like "distinctive hair" are impossible to grade. A button count isn't.
Generate the Same Character in Different Scenes
The candle test barely asked the model for anything. The room stayed the same, the pose stayed the same, and only one new object appeared. The real question is what happens when the room goes away entirely.
This is also the most common thing you will actually want. You want your OC standing somewhere else, doing something else, in different light. So we kept Ren as the reference and replaced everything around him.
We made one adjustment. We named the owl in each prompt so the model knew to include him, but we never described him. Where he sits came from the prompt, and what he looks like still had to come from the reference.
Here are the three prompts we ran, each setting a different place and a different lighting condition:
"he is standing at a night market stall with his owl on his shoulder, paper lanterns overhead, looking at something off frame"
"he is sitting on a hillside at sunset with his owl beside him, tall grass around him"
"he is walking through heavy rain holding an umbrella, his owl on his shoulder sheltering under it, wet cobblestones"
None of them said a word about Ren's hair, eyes, glasses, buttons, or embroidery.
Ren on the left next to three scene changes, where the eye colors, glasses, buttons, embroidery, and owl all carry over.
Most of the checklist came back intact. In particular, the owl is the headline. It came back as the same buff tufted bird in all three, whereas the prompt-only tests turned it into a different species every time.
Still, a scene swap asks very little. Ren stays exactly as he is in the reference and only the background and lighting get rebuilt around him. Changing his outfit or his pose means redrawing the character himself, which is a different problem.
Change Your OC's Outfit or Pose Without Losing Their Identity
A background swap is an easy ask for consistent character AI. Your character is left untouched while the world behind him gets rebuilt. An outfit or pose change is different, because the model has to redraw Ren himself and still return the same person.
So we tested that next. We kept the same reference image and ran three prompts, one that takes his cloak away entirely, one that keeps his clothes but changes how he is standing, and one that changes both at once.
Before looking at the results, it helps to be clear about what consistency actually means here. If you ask for a sweater, the cloak is supposed to disappear. The buttons and the embroidery go with it, and that is the request working rather than the reference failing. What should survive are the traits that belong to Ren rather than to his clothes, so the face, the hair, the heterochromia, the glasses, and the owl.
Here are the three prompts we ran:
"he is wearing a heavy knit sweater and scarf instead of his cloak, his owl on his shoulder, standing outdoors in the snow"
"he is crouching down to look at something on the ground, one hand resting on his knee, his owl on his shoulder, cobblestone street"
"he is wearing a summer yukata and sandals, sitting on wooden steps with one arm resting on his knee, his owl beside him"
None of them described his face, hair, eyes, or glasses.
Ren on the left next to a sweater, a crouching pose, and a yukata, with his eyes, hair, glasses, and owl carrying over into all three.
The identity traits held in all three. The amber eye stayed on the left, the dark eye on the right, the streak stayed on the correct side, and the glasses and the owl came back unchanged even in the yukata, which strips away the cloak, the buttons, and the embroidery at once.
The one thing that arrived unasked for is in the crouching image, where the model gave Ren an open mouth and chose a high camera angle. Neither appeared in the prompt. That is a useful pattern to notice, because when a prompt leaves a gap, the model fills it, and what it fills in is what will surprise you.
Testing Character Reference AI on a Difficult Angle
Every OC character generator test so far has shown Ren from roughly the same view. Front-on or close to it, face visible, both eyes in frame. That is the easiest thing to carry across, because the reference shows the model almost exactly what it needs.
An angle change takes that away. When the camera drops low or moves behind him, most of what the reference shows no longer lines up with the output, so the model has to infer the geometry and invent the parts it was never shown.
We ran three prompts, each hiding something different:
"low angle shot looking up at him, he is standing against an evening sky, his owl on his shoulder"
"seen from behind, he is walking away down a corridor and looking back over his shoulder, his owl on his shoulder"
"extreme close-up of his face turned three quarters away from the camera, his owl just visible behind him"
These were the results.
Ren on the left next to a low angle, a back view, and a close-up, where his eyes, hair, glasses, and owl carry over.
Our OC's identity survived the hardest reference image AI test. The amber eye stayed on the left and the dark eye on the right even in the close-up, where any slip would have been impossible to hide.
What the model did resist was the camera. We asked for a face turned three-quarters away and got one close to front-on, and the low angle came back milder than described. It kept pulling the composition back toward the view the reference gave it.
That trade-off is interesting. The further you move the camera from what the reference shows, the more the model has to choose between your composition and the design, and here it chose the design.
This is where Tsubaki.3's three reference slots earn their place. One image can only show the model one view, so a second or third from a different angle gives it something real to work from instead of a guess. If you know you need back views or dramatic angles, a front shot plus a side and a back is a stronger setup than one good portrait.
Character Reference AI for OC and VTuber Workflows
So far we've been running tests built to expose drift. But real work is more dynamic. You have one design that has to appear across a set of assets made weeks apart, and every one of them has to read as the same character.
For instance, if you draw an OC, that usually covers the following:
- Story illustrations: Your character moves through different locations across a scene or a chapter, and none of those locations should change who they are.
- Alternate outfits: A season, an occasion, or a full wardrobe set gives you several versions of the same person rather than several people.
- Seasonal artwork: A birthday post, a holiday piece, or an anniversary illustration reuses a design you already finished months ago.
- Scenes with other characters: Your OC has to hold up next to someone else's, where any drift is far easier to spot.
Similarly, if you run a VTuber channel, the assets are more promotional, but the underlying problem doesn't change:
- Announcement art and thumbnails: These need the avatar on-model at a glance, because a viewer scrolling past will recognize the design before they read anything.
- Outfit reveals: The clothes are the whole point of the post, so the face has to stay exactly where it was.
- Expression variations: Emotes, overlays, and reaction stills all show the same face doing different things, and they sit side by side where mismatches show.
- Seasonal assets: A stream schedule or an event graphic runs on a cycle, so the same character comes back several times a year.
One thing to keep in mind is that character reference AI in PixAI, including Tsubaki.3, produces flat images rather than a rig. It will not build or animate a Live2D model, and it doesn't replace the artist or rigger who does that work. What it covers is the promotional and static art around an avatar rather than the avatar itself.
What Else Can You Create From One Character Reference?
A reference isn't limited to producing just another illustration. Once a design exists, the same anchor works for most of the flat assets a character needs.
Here are some examples of artwork you can create using character reference AI:
- Expression sheets: A grid of the same face in different moods gives you emotes, overlays, or a reference page of your own.
- Character turnarounds: Front, side, and back views in one frame document a design the way an animation model sheet does.
- Manga panels: The character drops into sequential art without being redrawn for every panel.
- Stickers: A cropped set of poses and reactions becomes something you can actually use in chat or on a stream.
- Figure-style renders: The design gets treated as a physical object rather than a drawing, which is useful for merch mockups.
- Lighting and background variants: A composition you already like gets tested under different times of day or moods.
We picked one to test and created a three-view sheet or character turnaround using this prompt:
"Generate a three-view sheet of the @image1 character in the image."
This was our result.
A character turnaround built from our reference image shows Ren from the front, side, and back with his cloak, hair, glasses, and owl carried into all three views.
When to Use a Character Reference AI, a LoRA, or Image Editing
A character reference isn't the only way to keep a character consistent, and it isn't always the right one. There are three main approaches, and they solve different problems rather than competing for the same job.
The one we used in our tests works from an image you already have. There is nothing to train and nothing to set up beyond loading the file, which makes it the fastest route from a finished design to a new scene, outfit, or pose.
A LoRA takes the opposite approach. It stands for Low-Rank Adaptation, and it is a small add-on file that teaches a base model one specific character or style without retraining the whole model, which is the main difference when comparing a model and a LoRA.
You can train it on a set of images of your character, and from then on a trigger word brings them back without any reference image at all. That costs a training set and a wait, so it pays off when you expect to generate the same character over months rather than an afternoon. It also extends to multi-character LoRA setups, where two trained characters appear in the same scene.
Image editing is the third, and it is a repair tool rather than a generation strategy. When you already have the image you want and one thing is wrong, you change that part and keep the rest.
A comparison of character reference, LoRA, and image editing across what each one uses, what it is best for, and where it falls short.
The choice comes down to what you already have. If you have a design and need more images of it, reach for a reference. If you will need those images for the next year, train a LoRA. If you have the right image and one detail is wrong, edit it.
In fact, most people end up using all three. A reference gets you a set of clean images, those images become a LoRA training set, and editing cleans up whatever slips through afterward.
Tips for Getting Better Character Reference AI Results
When you start using Tsubaki.3 with your own character, a few tips can make things much easier. For example, a smooth workflow starts with the reference itself.
Our library shot works because nothing sits in shadow and every trait is visible at once, and the tests that held best were the ones where the output view stayed close to what that image already showed.
From there, describe only what changes. Every scene prompt said nothing about Ren's hair, eyes, or clothes, and they came back correct anyway, because re-describing the character competes with the reference instead of supporting it.
That same logic applies to anything traveling with him. We told the model the owl belonged on his shoulder and never described the bird, and it came back as the same owl every time. In the prompt-only tests, where we described it carefully and had no reference, it changed species in every single image.
Expect some pushback on composition, though. When we asked for a low angle, a back view, and a three-quarter turn, all three came out closer to the front-facing reference than the prompt described.
The model protects the design at the expense of your framing, so the further your camera moves from the reference view, the less that instruction lands. If you need those angles, load a second or third reference rather than fighting it, since Tsubaki.3 takes up to three.
Finally, check your results as a set rather than one at a time. Small drift is invisible in isolation and obvious in a row.
How Consistent Is Tsubaki.3 for Character Reference AI?
Across roughly a dozen generations from one reference image, Ren stayed recognizably Ren every time. But recognizable isn't identical, and the gap has interesting insights.
His face held best. The shape, the proportions, and the eyes carried through scene changes, outfit swaps, a crouch, a low angle, and a back view.
The heterochromia stayed consistent in every single-image test, the glasses kept their frames, and the owl came back as the same bird even though we never described it. The clothing was steadier than expected too, with three buttons staying three and the embroidery surviving low light, heavy rain, and a shot from below.
What moved was smaller and more scattered. The rain tightened his cloak into a fitted overcoat, and the crouch added an open mouth and a high camera angle that appeared nowhere in the prompt. Parts the reference never showed, like his legs and the back of the cloak, were invented rather than carried over.
Three details changed without being asked for, including a fitted coat in the rain, an open mouth in the crouch, and invented trousers below the back of the cloak.
So Tsubaki.3 holds a character well enough to build a real set of assets from one image, as long as you check the results as a group and keep an eye on the small details. It is a strong anchor rather than a lock.
Your OC Already Exists, So Stop Rebuilding Them
Anime studios solved this problem decades ago by handing every animator the same model sheet. Nobody redraws the character from memory, because nobody has to. The design already exists, so the job is to work from it.
That is the whole idea behind character reference AI. Once you have an image of your OC that came out right, describing them again from scratch is redoing work you already finished, and every attempt is another chance to lose something.
But it isn't perfect. A reference is an anchor, not a lock, and small details still need checking.
If you have a character sitting in a folder somewhere, that image is all you need to start. Load it as a reference with Tsubaki.3 in PixAI, pick a scene you have never drawn them in, and see where they turn up.
Top comments (0)