Debugging Character Drift: A Controlled Anime Character Creator Test

Which anime character creator keeps your OC the same after the first image? Explore my hands-on test of PixAI, NovelAI, and Leonardo.Ai across five tests.

If you’re building a story, a comic, or a VTuber concept with AI, one great portrait isn’t enough. You need the same original character (OC) in dozens of images, and that’s where many AI generators start to drift.

In fact, that’s the real test of an anime character creator. If you only need a single avatar, a preset avatar maker will do the job. But keeping the same face, hair, and signature details across new expressions, outfits, poses, and scenes is much harder.

So I wanted to find a tool that could keep my OC looking like herself, no matter what I changed. To find out, I set up a small side-by-side test.

I created an OC named Suzune and ran her through the same five tests in PixAI, NovelAI, and Leonardo.Ai. It involved four changes and one targeted fix in each tool.

Suzune’s master portrait next to four PixAI Tsubaki.3 results from my anime character creator test.

So which one held up? PixAI was the only tool with a usable result in all five tests. Leonardo.Ai’s free plan made the cleanest big changes but couldn’t fix a small detail. NovelAI kept Suzune’s look tightly but often underdid the changes I asked for.

I tested everything myself on Nov 3, 2026, using just one character, so you can think of this as one creator’s experience rather than a final verdict. This article is also in collaboration with PixAI, and I earn referral credits when someone joins through my link below. Now let’s dive right in!

How to Create Your First OC Portrait in PixAI

Every test started from the master reference below. Suzune is a shrine keeper’s apprentice who spends her days off in a café. I gave her a low side ponytail on her right and one gold bell on her left wrist, so it was easy to spot when a tool flipped or forgot a detail.

My master reference for Suzune, made in PixAI with Tsubaki.3 on the first attempt.

You can recreate this portrait or make your own OC by following the four steps below.

Step 1: Copy the Tested Anime OC Prompt

The prompt below gave me the portrait above on the first try with Tsubaki.3, PixAI’s newest anime model. I didn’t need a LoRA or Recipe, so you can use it as is.

1girl, solo, young adult woman, full body, standing, facing viewer, arms relaxed at sides, both hands visible, neutral expression, soft heart-shaped face, large round green eyes, long straight dark crimson hair, low side ponytail tied on her right side, wispy side bangs, small gold bell on a red cord around her left wrist, cropped white jacket with wide sleeves and pale pink inner lining, dark green pleated hakama-style skirt, plain light gray background, clean anime illustration, soft cel shading, crisp lineart

Next, add a negative prompt. This tells the model what to avoid. I used it to block the changes I was most worried about, like a red shrine-maiden skirt, twin tails, or extra bells.

red hakama, twin tails, centered ponytail, hair bell, hair ornament, multiple bells, hat, glasses, text, watermark, extra fingers, blurry

Step 2: Customize Your OC’s Hair, Eyes, Outfit, and Accessory

Here are the phrases to swap in the prompt to make the character your own:

• Hair: long straight dark crimson hair, low side ponytail tied on her right side
• Eyes: large round green eyes
• Outfit: cropped white jacket with wide sleeves and pale pink inner lining, dark green pleated hakama-style skirt
• Accessory: small gold bell on a red cord around her left wrist

Keep one small accessory on one side only, since that’s often the first detail to drift.

Step 3: Set Up Tsubaki.3 and Generate Your Portrait

Once your prompt is ready, follow these steps to generate your portrait with the same settings I used:

1. Open Image Generation in PixAI’s left menu.
2. In the Model tab, choose Tsubaki.3.
3. Switch to the Settings tab. Under Number of images, choose Single, so you get one image at a time instead of a batch of four. Set the Aspect Ratio to 3:5, a tall portrait shape that fits a full-body character, and the Resolution to M (768×1280). I left the Quality Tag booster off so only the prompt controlled the result.
4. Paste your prompt and negative prompt, then click Generate.

The exact setup I used in PixAI’s image generator, with a 3:5 ratio, M size, single image, Turbo Mode, and a 3,000-credit cost shown before generating.

Step 4: Check the Credit Cost Before You Start

On my Starter plan, this setup cost 3,000 credits with Turbo Mode at +0 credits, and one attempt was enough. Free accounts can generate at Low Priority, also at +0 credits.

Ready to make your own? Create your first OC portrait in PixAI with Tsubaki.3 using the prompt above. Swap in your character’s details, then save the result as your master reference.

If you’re new to PixAI, you can join through my invite link. Once you verify your email (email sign-ups only), create at least one image, and claim the Daily Reward on two separate days, we each get 50,000 credits within 24 hours. The bonus comes after those steps, so start with the credits already in your account. And yes, I get the same reward when you join this way.

How I Tested Three AI Tools for Anime Character Design

To keep the comparison fair, I used the same PixAI master image as the reference in all three tools. Every change started from that original image, so mistakes couldn’t build up from one edit to the next.

For each tool, I followed its own documented reference workflow and wrote prompts in the style its docs recommend.

Here’s the setup I used in each tool:

The three AI tools for anime character design in my test, with the plan, model, reference route, and output size I used for each.

To keep things fair, I saved the first result from every task, even the bad ones. If something went wrong, I gave each tool one more try and changed only one thing in the prompt or settings. I didn’t touch up any image outside the tools.

The free plans did run out partway through. NovelAI’s free trial covers 30 images, and Leonardo gives you 150 tokens a day, so I finished both on a second free account with the same settings.

When judging each result, I looked at two things separately. First, did the tool actually make the change I asked for? Second, does Suzune still look like Suzune? For that, I checked her face, green eyes, crimson hair, which side her ponytail is on, the single wrist bell, and the clean anime style.

Can a Consistent Anime Character Survive Four Changes?

To test each tool the way you’d actually use it, I gave Suzune four everyday changes. She got a surprised expression, a casual outfit, a seated pose, and a trip to a café, one change at a time.

I’ve included the PixAI prompt for each test so you can adapt it for your own OC. Each one follows a simple rule from PixAI’s Tsubaki.3 prompt guide. You describe only the change, then name the details you want to keep.

For NovelAI and Leonardo.Ai, I asked for the same change but wrote each prompt the way that tool’s docs recommend, using tags for NovelAI and plain editing instructions for Leonardo.

Test 1: Changing Your OC’s Expression Without Losing Her Face

Here’s the PixAI prompt I used to make her visibly surprised:

The character from @image1, visibly surprised, eyes wide, eyebrows raised, and mouth slightly open. Keep everything else unchanged, including the low side ponytail on her right side and the small gold bell on a red cord around her left wrist.

Only PixAI and Leonardo made the surprise clear.

All three tools kept Suzune recognizable, but they handled the surprise very differently. PixAI changed only her face, and even her hands and skirt hem still match the master.

Meanwhile, Leonardo gave the strongest surprised expression, with raised, slightly knitted eyebrows. And NovelAI kept every detail but barely moved her eyes or brows. Its docs warn that a high Strength setting can make expressions look too much like the reference, and I’d left it at the default of 1.

Test 2: Swapping the Outfit While Keeping Signature Accessories

Here’s the PixAI prompt I used for the casual outfit:

Dress the character from @image1 in an oversized cream knit sweater and a long brown skirt, keeping her face, hairstyle, hair color, the low side ponytail on her right side, and the small gold bell on a red cord around her left wrist exactly as in the reference.

PixAI and Leonardo swapped the whole outfit and kept the wrist bell. NovelAI changed the sweater but kept a pleated skirt that fades from green to brown.

Two out of three tools nailed the full outfit swap. PixAI and Leonardo replaced both the jacket and skirt, and Leonardo even added a cable-knit texture to the sweater.

Similarly, NovelAI changed the top. But the skirt stayed pleated and faded from olive green to brown, so part of the old outfit leaked through.

The bell surprised me most. I expected it to vanish with the wide sleeves, but it survived in all three tools.

Test 3: Changing the Pose From Standing to Seated

Changing the pose was PixAI’s hardest test. The first attempt seated her correctly but turned her cropped jacket into a long wrap top.

For the retry, I named the exact items to keep:

The character from @image1, sitting on a plain wooden stool with her hands resting on her lap and her left wrist visible. Keep her appearance and outfit unchanged, including the cropped open white jacket with wide sleeves, the small gold bell on a red cord around her left wrist, and her black boots.

Every tool got her seated. PixAI’s first attempt redesigned the jacket, and the retry restored it.

Naming the exact items to keep made a big difference here. PixAI’s retry brought back the cropped jacket, though it added a collar, a chest pocket, and heeled boots.

NovelAI’s first attempt closed her eyes in a smile and gave her two bells. Naming the expression with “neutral expression, looking at viewer” fixed her face, but the second bell stayed. Leonardo was the only tool that kept every detail on the first try.

Test 4: Placing Your Character in a New Scene

Here’s the PixAI prompt I used to move her into a café:

Put the character from @image1 inside a cozy café, standing beside a wooden table, with warm light coming through the window behind her. Keep the character's appearance exactly as in the reference.

Leonardo kept her exact pose and framing in the café. PixAI’s version turned more painterly, and NovelAI cropped the frame and changed her pose.

All three placed her in a café. Leonardo kept her exact pose and framing because its docs recommend asking for that directly in the prompt, and I did. PixAI’s scene turned more painterly, and her hair shifted toward magenta in the warm light. NovelAI cropped her at the thighs, raised her arms to hold a cup, and added a second bell again.

If your OC holds up through changes like these, you’re ready for a bigger project. You can turn your best results into a character reference sheet that keeps all her key views and details in one place, which is useful for comics, commissions, or a future VTuber model.

What Happens When One Detail Needs Fixing?

Sometimes a result is almost perfect except for one small mistake. Regenerating the whole image can fix it, but you also risk losing everything that was already right.

Masked editing can help. You can paint over just the problem area, describe what should be there instead, and the tool redraws only that part.

NovelAI’s doubled bell from the pose test gave me the perfect flaw to try this on. It’s small, easy to spot, and sits on Suzune’s signature accessory. To keep the test fair, I took that same seated image into all three tools, masked the bow, cord, and both bells, and asked for a single bell.

The same flawed image went into each tool to turn two bells into one.

PixAI’s Repaint tool fixed it on the first attempt. I used Masked Area with the default settings (Feathering 8, Strength 0.75, Default mode) and this prompt:

A single small gold bell hanging from the red cord bow on her left wrist.

It redrew a crisp bell and bow, and the hand and skirt near the mask edge stayed the same. The fix cost 1,500 credits, less than half the 3,900 credits of a Tsubaki.3 edit.

PixAI Repaint with the bell area masked. I kept the default Feathering 8, Strength 0.75, and Default mode, and the fix cost 1,500 credits.

NovelAI’s Inpaint tool also fixed it on the first attempt. I kept Precise Reference on with the master image (Character & Style, Strength 0.8, Fidelity 1) and left Inpainting Strength at its default of 1. Here’s the prompt I used:

1girl, jingle bell, single bell, wrist ribbon, red ribbon

In Undesired Content, I added “multiple bells, twin tails, hair ornament, text, watermark”. The new bell looks slightly softer than PixAI’s, and nothing outside the mask changed.

Leonardo.Ai’s Canvas editor was a different story. On the free plan, it ran Leonardo Diffusion XL, an older SDXL model. My first attempt used this prompt:

There are two gold bells hanging from the red cord on the woman's left wrist. Change this to a single small gold bell. Keep everything else unchanged, including her face, hands, sleeves, skirt, pose, framing, and background.

That gave a metal-looking bangle and no bell. For the second attempt, I switched to a shorter prompt, “a single small gold bell hanging from a red cord bow”, but the bell still didn’t appear.

Leonardo’s docs say turning on Outpaint disables Inpaint Strength, and Outpaint was on during my first two attempts. So I ran one more troubleshooting attempt with Outpaint off, Inpaint Strength at 0.6, and a more familiar description:

a red string bracelet tied in a small bow, with a single small gold bell charm, anime illustration

I also added “bangle, metal bracelet, gold bracelet, multiple bells, chain” as the negative prompt. That gave a red knot instead of a bell. Finally, I tried an unmasked edit with Nano Banana using the same first-attempt prompt, and it removed the bell and cord entirely.

So having a mask wasn’t enough on its own. The editor also had to understand a small anime detail like this one.

Which Anime Character Creator Fits Your Project?

After five tests, each tool had clear strengths and weak spots. Here’s everything side by side, from how well each one kept Suzune’s identity to what it cost me. Each tool uses its own credits or tokens, so the costs aren’t comparable as money.

How PixAI, NovelAI, and Leonardo.Ai compared as anime character creators in my Suzune test. PixAI was the only tool with a usable result in all five tasks.

So which one should you pick? It depends on where you are with your character. Here’s how I’d choose based on what you’re making:

• If you’re starting a new OC, go with PixAI. You can make your master with the prompt above, build variations from it, and fix small details in Repaint without switching tools.
• If you’re expanding an existing design on a free plan, Leonardo.Ai handles big outfit, pose, and scene changes cleanly. Just plan to fix small details elsewhere, and keep in mind that free generations show up in the public feed.
• If you want tight control over the reference, NovelAI suits creators who don’t mind adjusting Strength and Fidelity for each task.

If I were continuing with Suzune, I’d stay in PixAI. It was the only tool where I could create her, change her, and fix her without switching apps. If you want to try the same workflow, start with the prompt in Step 1 and make your first OC portrait in PixAI.

Anime OC FAQ

Here are some questions that came up while I was testing:

• How do I keep the same anime character across different images? Use one clear master as the reference every time, describe only the change, and name the details to keep. Edit from the master, not your last edit.
• Do I need to train a LoRA before making OC variations? No. Every result here came from one reference image. A LoRA helps later, when you need many images in the same style.
• Why does my character change when I change the outfit? The tool redraws a large area, so nearby details like sleeves and wrist accessories shift too. Name accessories in the prompt, and fix any that drift with a masked edit.
• Can I use these illustrations as a VTuber model? Not directly. A Live2D model needs the design split into layers, such as hair, eyes, and mouth, and then rigged. These images work as the design reference for that.