Consistent Anime Characters from Reference Images: Testing Tsubaki.3's Character Reference

Making one great anime character is the easy part. Making that same character again is where it falls apart.

You generate an OC you love, then the next image comes back with a different face, a shifted hairstyle, the wrong eye colour, altered body proportions. Change the pose or camera angle and the gaps widen. For stories, comics, VTuber assets or any recurring illustration, that drift turns into a real problem fast.

Character reference AI is the answer to this. Rather than making the model rebuild your character from a text description each time, you hand it an existing image to work from visually.

I put that approach through Tsubaki.3 on PixAI, running the same OC reference across three separate generations. One question drove it: how far can a single existing anime character design carry into new scenes, outfits, poses and lighting?

Why Is It Hard to Generate the Same Anime Character Consistently?

A prompt can describe a character's major traits well enough — hair colour, eye colour, hairstyle, clothing, accessories, whatever identifies them.

What a prompt can't do is remove interpretation.

Describe a character as having long silver hair, red eyes, a navy ribbon and a black-and-white outfit, and nothing guarantees each generation renders those details identically. The face shifts slightly. The hairstyle moves. An accessory vanishes.

It gets harder as you change the:

• Scene
• Pose
• Camera angle
• Outfit
• Expression
• Lighting

More change means more room for the character to drift.

That's the difference between using an OC character generator for a series versus one standalone image. Comics and stories need readers recognising the same character panel to panel. VTuber creators need promotional artwork, alternate outfits, seasonal illustrations and other assets built on one established design.

Prompt-Only Generation vs. Character Reference

Telling an AI what a character should look like and showing it what that character already looks like are different operations.

Prompt-only, you'd write something like:

anime girl, long silver hair, red eyes, navy ribbon, black and white dress, detailed anime illustration

The model interprets and builds from that. Repeat the prompt and broad traits survive, but the smaller details wander.

A reference image AI workflow gives the model a second source. Instead of reconstructing the character purely from words, it anchors to an existing image.

Put simply:

A prompt describes what the character should look like. A reference shows what the character already looks like.

Which makes character reference AI genuinely useful for anyone reusing an established design without rebuilding its appearance from text every time.

Perfect consistency isn't on offer. The new image is still generated, so details can shift. But a consistent character AI approach hands the model visual information no text description can carry.

How Character Reference AI Works in Tsubaki.3

I used one original OC as the main reference throughout, to see how well character reference AI holds a design across different generations.

Tsubaki.3 accepts up to three Character Reference images, drawn from previous PixAI generations or your own uploads. The workflow:

1. Start with an existing character image.
2. Select it as the Character Reference.
3. Choose Tsubaki.3.
4. Write a prompt describing what you want to change.
5. Generate the image.
6. Compare the result directly with the original reference.

That last step is the one people skip. A generated image can look great on its own and still fail a consistency test. Attractive isn't the question — recognisable is.

Across the tests below I held the reference constant and varied different aspects of the scene.

Generate the Same Character in Different Scene

The gentlest way to generate same character in different scenes: leave the character mostly alone and relocate her.

Test 1: Café Scene

Reference:

Prompt:

@image1 sitting at a cafe table, drinking a warm cup of coffee, indoor cafe background, ambient warm lighting, rainy window in background, night time, cinematic lighting, masterpiece, detailed anime artwork

Setting, lighting, activity and pose all change; character design goes largely unspecified.

Generated Result:

Evaluation:
All four outputs place her convincingly in the café. More to the point, she stays highly recognisable.

Silver hair, red eyes, navy ribbon and the black-and-white outfit all carry through consistently. The face holds very close across outputs, with expressions and facial positioning varying a little. Hair shape moves naturally with the different compositions while the core hairstyle survives.

A strong showing for consistent character AI: environment and pose change without the character becoming someone else.

Change the Character's Outfit or Pose Without Losing Their Identity

Swapping the background is a soft test. Changing clothing and pose is harder, because it alters the character rather than the surroundings.

Here a consistent anime character generator has to hold identity while still executing the requested changes.

Test 2: New Outfit and Pose

Reference:

Prompt:

@image1 dynamic pose, standing, full body, wearing an oversized beige knitted sweater, denim skirt, black tights, winter outdoor park background, falling snow, holding a hot drink, cheerful subtle smile, masterpiece, highly detailed

This one changes nearly everything around the original design: outfit, pose, setting, weather, activity, expression.

Generated Result:

Evaluation:
The requested changes land clearly. She's standing in a snowy park now, beige sweater, denim skirt, black tights, drink in hand.

The major outfit change notwithstanding, she stays easy to identify. Silver hair, red eyes, navy ribbon and general facial design all carry over. Full-body proportions hold reasonably steady across the outputs despite the varied poses.

Which surfaces something important: character consistency doesn't mean keeping every visual detail identical.

Outfit and pose are supposed to change. Identity is what should hold.

Test Another Variation: Side Profile and New Lighting

The third test applies a different kind of pressure. Rather than the outfit, I changed camera angle and lighting.

Test 3: Side Profile at Sunset

Reference:

Prompt:

@image1 close-up face portrait, side profile, thoughtful gaze, sunset lighting, golden hour glow, dramatic rim lighting, soft shadow, detailed face, masterpiece, cinematic angle

Framing shifts from the original front-facing portrait to a close-up side profile, with strong golden-hour lighting introduced.

Generated Result:

Evaluation:
She stays recognisable through the harder angle.

Silver hair, the visible red eye, navy ribbon, white collar and black outfit all hold across outputs. Side-profile framing naturally changes how the face reads, but her defining features persist.

Lighting also shifts the apparent colour and shading of hair and skin. Expected, not a consistency failure. What counts is that the underlying character design survives the dramatic lighting change.

Across all three, Character Reference supports substantial changes to clothing, pose, composition and lighting without losing the original character.

Character Reference for OC and VTuber Workflows

For OC creators, this makes a character design far more reusable.

One reference becomes a starting point for:

• Story illustrations
• Different scenes
• Character interactions
• Alternate outfits
• Seasonal artwork

VTuber creators working with an AI character reference generator may want promotional illustrations, alternate outfits, expression visuals, seasonal assets or scene illustrations that keep the character recognisable.

None of this replaces Live2D rigging or a full VTuber production pipeline. It's a way to build more visual assets around a design you've already established.

Rather than starting from zero each time, start from the character you already like.

What Else Can You Create from One Character Reference?

Character Reference also extends an existing design into other asset types:

• Expression sheets
• Character turnarounds
• Manga-style images
• Stickers
• Different lighting or background setups
• Figure-style images

The point isn't using every available format. It's building on an existing design instead of reconstructing it from text over and over.

For an OC creator, that turns one successful illustration into the foundation for a whole collection.

Character Reference vs. LoRA vs. Image Editing

Useful as it is, Character Reference isn't the right answer to every consistency problem.

Character Reference

Best when you already hold one or more character images and want new scenes, poses or outfits quickly.

No training required first. You supply the visual reference at generation time, which makes it a practical AI character reference generator workflow for extending an existing OC.

LoRA

Better suited to a trained, reusable character workflow — worth it for creators expecting to generate the same character repeatedly across a longer project.

PixAI's Model vs. LoRA: Understanding PixAI Foundations covers the distinction. For multiple recurring characters, there's also a Multi-Character LoRA Generation Guide.

Image Editing or Inpainting

Better where you want to preserve one existing image and alter only a specific part.

Changing an outfit while keeping the original composition, background and character placement, for instance, is an editing job.

In short:

Different problems, different tools. Character Reference doesn't universally replace LoRA or editing.

Tips for Better Character Reference Results

A few practical lessons came out of the Tsubaki.3 tests for anyone setting up a consistent character AI workflow:

Choose a clear reference

Pick an image where the identifying features are clearly visible. The reference used here shows hair, eyes, ribbon, face and outfit plainly.

Be specific about what should change

These prompts worked because they described the new scene, outfit, pose or lighting rather than trying to redesign the character.

The café test concentrated on environment and activity; the second specified new clothing and pose.

Avoid unnecessary redesign instructions

Where the reference already establishes the character, repeating every trait in the prompt isn't necessary. Focus on what you actually want changed.

Compare multiple outputs

One image won't tell you about consistency. Several results make it far easier to see which traits hold and which drift.

Use additional references when they genuinely help

Tsubaki.3 takes up to three Character Reference images. Multiple references help where one image doesn't clearly show an important part of the character.

More isn't automatically better, though. Useful visual information is the goal, not image count.

How Consistent Is Tsubaki.3 in Real Use?

On these three tests, Tsubaki.3 performed well as a consistent anime character generator — holding overall identity while permitting substantial visual change.

Most consistent traits:

• Silver hair
• Red eyes
• Navy ribbon
• General face and hairstyle
• Black-and-white character design
• Overall visual identity

All recognisable across the café, winter outfit and side-profile tests.

Variation between outputs remained. Facial positioning and expressions changed; hair and face shifted naturally with pose, angle and lighting.

Most visible in the side-profile test, where the face has to be reconstructed from a fundamentally different viewpoint.

So this isn't 100% consistency or pixel-perfect preservation.

A better test: would you recognise the outputs as the same character?

Here, yes.

Final Thoughts

If you've already got an anime OC you love, rebuilding that character from text every time isn't the most practical route.

Character reference AI offers another: begin with the character you have and use that image as a visual anchor.

The Tsubaki.3 tests show one reference carrying into new:

• Scenes
• Outfits
• Poses
• Lighting setups

Treat consistency as preserving identity, not freezing pixels. What you ask to change should change. What matters is whether the defining features stay recognisable.

For quick OC illustrations and recurring character assets, Character Reference extends an existing design rather than rebuilding it each time. For a longer-term trained workflow, a LoRA fits better. To preserve a specific composition while changing one element, image editing is the better call.

Got a character design you want to reuse? Try it with Tsubaki.3 on PixAI and compare every generation directly against your original reference.