How to Create Consistent Anime Characters from Reference Images

For the certified otakus reading this, have you ever created an original anime character and then wanted to see that same character in a different pose, outfit, or setting?

Some of my friends have tried doing this with ChatGPT and Gemini. They can produce good anime images, but the results are not always the version of the character they had in mind.

As an AI enthusiast, I went looking for a tool built more specifically for this use case. One that I recommend trying is PixAI, an AI creative platform focused on anime art. It includes image generation and editing tools, anime models, community LoRAs, character references, and other workflows for developing original characters.

PixAI recently introduced Tsubaki.3, its newest anime image model. It can take a character reference image, identify important visual traits, and reinterpret the subject across new poses, outfits, camera angles, and scenes. Tsubaki.3 is still in early access at the time of writing, so it may not yet appear on every account.

In this guide, I will show how the new reference workflow works, including single-character edits, multi-character replacement, pose transfer, and turning a 2D character into a figure-style render.

Let’s get started.

What’s New in PixAI?

Let me start with PixAI’s redesigned image generator dashboard. Reference images can now be dropped directly into the prompt area instead of being managed through a separate part of the interface. I find this layout much easier to follow because the images and instructions are kept in one place.

PixAI’s image generator dashboard with Tsubaki.3 model. Image by Jim Clyde Monge

The new dashboard also felt faster and more reliable during my tests. After uploading an image, I could refer to it inside the prompt by typing @ and choosing the corresponding file. This becomes especially useful when working with several references because I can tell PixAI exactly which image contains the character, outfit, pose, or base composition.

PixAI’s image generator dashboard with Tsubaki.3 model. Image by Jim Clyde Monge

The older generator has not disappeared. Clicking Switch to Classic Editor returns to the previous dashboard for anyone who prefers that workflow or needs an option that has not yet moved to the new interface.

The other major addition is Tsubaki.3. PixAI designed the model to support more of the anime creation process rather than producing unrelated one-off images. According to the official Tsubaki.3 showcase, one reference can be enough for the model to extract a character’s traits and place that character in a new scene.

Tsubaki.3 model. Image by Jim Clyde Monge

Its key capabilities include:

• Reference-based character generation: Upload an existing character and reuse their identity without training a separate model first.
• Character and style consistency: Preserve recognizable facial features, hair, accessories, and the broader visual style across different outputs.
• Outfit, pose, and scene control: Combine a character from one image with clothing, body positioning, or an environment taken from another.
• Single-character and multi-character compositions: Give each character a separate reference and identify them individually inside the prompt.
• Creative production formats: Generate expression sheets, turnarounds, manga panels, stickers, alternate lighting setups, and figure-style renders.

This turns Tsubaki.3 into more than an AI character reference generator. It can also work as a consistent anime character generator for creators who need the same subject to survive several visual changes without being redesigned every time.

Okay, now let’s move on to the new features that you can play with.

Maintain Character Features Across New Scenes, Outfits, and Compositions

The first workflow I tested was an outfit transfer. I wanted to keep the identity of the character in the first image but dress that character in the clothing shown in the second image.

Here are the reference images and the prompt I used.

Prompt: Keep the head features of the character in Image 1 unchanged, and only replace their clothing with the clothing from Image 2.

Sample input images in manipulating character pose with a reference image. Image by Jim Clyde Monge

Upload these two images along with the text prompt on the image generator tool. This is what your dashboard should look like:

Sample input images in manipulating character pose with a reference image. Image by Jim Clyde Monge

Notice how I am tagging the reference images; just type “@” and select the reference image. Submit and wait for the final image to be generated.

Here are the two sample versions:

I like that the two versions do not feel like duplicates. One stays very close to the original character, while the other allows more of the second reference’s styling to influence the result. Both still preserve the features I cared about instead of producing a completely new person wearing a similar outfit.

This is useful when developing alternate costumes for an original character. A creator could start with one approved design and test a school uniform, formal outfit, battle costume, or seasonal variation without rewriting the entire character prompt. The reference carries the identity while the instruction explains what is allowed to change.

Single-character and multi-character references

Keeping one character consistent is already difficult, but a scene with two characters introduces another problem. Their traits can blend together.

Hair colors move from one subject to another, outfits become mixed, or the model changes both characters even though only one was supposed to be replaced.

Tsubaki.3 lets me attach separate reference images and assign each one a specific role. In the example below, I used a two-character image as the base composition, then uploaded another character to replace only the person on the left.

Here’s an example:

Manipulating character pose with a reference image results. Image by Jim Clyde Monge

This is the sample shortened prompt used:

Prompt: “Using as the base, perform a character replacement. Preserve the overall composition, camera angle, urban street background, sunny natural lighting, and the close, friendly photo-taking interaction between the characters in . Keep the character on the right completely unchanged, and only replace the character on the left side of with the character from”

The replacement character sits naturally inside the original composition. The camera angle, street background, lighting, and interaction between the two subjects remain close to the base image.

More importantly, the character on the right was not redesigned while PixAI worked on the left side.

The replacement also retained the details that made the source character recognizable, including the face, dress, eyepatch, hair, and accessories.

This is the kind of result I would want from consistent character AI because it follows the requested division of work. One character changes while the rest of the scene remains stable.

Being able to tag each reference separately also makes multi-character projects easier to organize. I do not have to generate every subject in the same conversation or combine them manually before uploading.

I can provide one clean image for each character, use another image for the composition, and state which person should appear in each position.

Using a Separate Pose Reference

The same tagging system can be used for poses. I can upload an already finished character illustration as @image1, then attach a sketch, photograph, or 3D mannequin as @image2. The second image provides the body position without forcing me to copy its identity, outfit, or background.

A prompt can be as direct as this:

Prompt: Keep the identity, hairstyle, facial features, and outfit of the character in @image1. Place the character in the pose shown in @image2. Use @image2 only as a pose reference and do not copy its appearance or background.

For difficult camera angles, I would add the framing and perspective in text, such as “full-body low-angle shot” or “side profile from waist level.” A pose reference gives the model the physical arrangement, while the prompt clears up anything the image leaves uncertain.

This is useful for action scenes, character sheets, thumbnails, and manga panels. It also reduces the need to describe every arm and leg position through a long prompt, which can become confusing once the pose gets complicated.

Turning a 2D Character Into a Figure

Character references can also change the presentation format. In the example below, PixAI turned a flat anime character into a PVC figure-style render while retaining the original design.

Turning a 2D image into 3D with Tsubaki.3. Image by Jim Clyde Monge

This is the prompt used for the example above:

Prompt: “Generate a PVC figure of this @image1 character, Glossy painted plastic surface with smooth highlights”

This does not create a printable 3D model file by itself. It creates a visual concept of how the character could look as a physical figure. An artist or 3D modeller could then use the result as reference when preparing the actual geometry, materials, pose, and display base for printing.

Where OC and VTuber Creators Can Use This

For original character and VTuber projects, one finished portrait is rarely enough. A usable character library may need front, side, and back views, several facial expressions, alternate outfits, promotional illustrations, reaction images, thumbnails, and scenes involving other characters.

Reference generation gives creators a faster way to expand that library. An OC character generator can take an approved design and explore how the character looks in a café, on a stage, during a fight scene, or in a different season.

VTuber creators can use the same process to plan costumes, expression sheets, stream graphics, and merchandise concepts before commissioning or drawing the final assets.

I would not rely on one front-facing portrait for every possible angle. The model cannot reliably infer details that were never visible, such as the back of a jacket or the exact structure of a hairstyle from behind.

For a long-running project, I would gradually build a small set of approved references from several angles.

PixAI’s OC and VTuber character guide also recommends combining reference images with stable prompts and, when stronger long-term consistency is needed, a trained character LoRA.

There are a lot more things you can do with PixAI’s new image editor plus its brand new image model, Tsubaki.3. I encourage you to explore the various use cases that you can do with it.

Tsubaki.3 Showcase | PixAISee what Tsubaki.3 can do across nine capabilities - coloring breakdowns, character consistency, lighting, manga…

Final Thoughts

So there you have it, you just read how Tsubaki.3 gave me a practical way to generate the same character in different scenes without training a custom model first.

As you’ve seen in the examples, I could transfer an outfit, replace one person inside a two-character composition, use a separate pose reference, and reinterpret a 2D design as a figure while keeping the subject recognizable.

A huge props to platforms like PixAI for giving anime creators access to models developed around their actual workflows. Character consistency is needed for comics, VTuber assets, visual novels, game concepts, and any project that uses the same cast more than once. Tsubaki.3 makes those experiments easier to start, even for someone who has not trained a LoRA or used a traditional image editor.

Tsubaki.3 remains in early access, but the official showcase already includes more examples involving character references, outfits, poses, manga, expression sheets, and multi-character scenes.

Check out PixAI and let me know what you think!