Published: Oct 22, 2026Last updated: Oct 22, 2026

Which AI is best for making PNGTuber assets?

31 min read

From character design to main art and variant images, I prepare PNGTuber assets with several image AIs and work out which one is best for the job.

Images of a blonde girl in a military-style dress made by five AIs, in 5 columns and 2 rows. The top row is character design, the bottom row is main art regenerated from a reference image. From the left, ChatGPT, Gemini, Grok, PixAI, and ComfyUI

I’ve had all sorts of AIs generate images while making samples for Peraravatar+, but which one is actually best for a PNGTuber? I had a rough answer in my head, but I had never checked it, so I decided to have several image AIs output avatar images again and find out which AI is best for making PNGTuber assets.

As it happens, a streamer friend’s birthday is coming up, so I’ll use their REALITY avatar as the reference and try out different ways of generating while preparing PNGTuber avatar assets for Peraravatar+. REALITY is a streaming app for phones where an avatar moves along with your face on the front camera. Many people who move from REALITY to YouTube and other platforms make a new avatar based on the design they were using, so I’ll use that as the example.

Our test subject: Piyochama’s REALITY avatar

When I asked, “Can I use your REALITY avatar as a reference for a PNGTuber note?”, they happily said yes. Yet when I once suggested, “Let’s get a domain and do some SEO so your official page shows up in search!”, they turned it down because it was too embarrassing, so I won’t link to them directly. The official site I made for them is in Achievements.

What I compared

I had ChatGPT, Gemini, Grok, PixAI, and ComfyUI output images from similar prompts and compared them. Everything is a free service or tool, and I don’t use any paid-plan options. Here’s what I compared.

Character design: With no concrete image in mind, does it output a nice character from just a rough description?
Regenerating from an image: Given a reference image (or your favorite from the character design if you don’t have one), does it output main art for an avatar?
Variant images: Given one image picked from the regenerated ones, does it output the same image with only the eyes closed and everything else unchanged?

Many AIs are bad at outputting images with a transparent background, so I use a white background for all the main art.

Prompts and screenshots: I wrote every prompt in Japanese. The prompt boxes in this English version show translations, except for PixAI’s Japanese prompts, where the Japanese itself is the point. The screenshots are from the Japanese UI.

What people generally say: General comparisons say ChatGPT follows instructions well and is good at text in images, Gemini is good at keeping a character consistent, and Grok is good at photo-like images but opinions are split on its anime style.

Rights to generated images: Under their terms of use, ChatGPT, Gemini, and Grok let you use generated images commercially (as of October 2026). With PixAI and ComfyUI, whether you can use them commercially depends on the model and LoRA you use.

Turning off training on your data

By default, ChatGPT, Gemini, and Grok use your uploaded images and conversations to train their AI. This time I’m handing over someone else’s avatar, so I turn that off first.

ChatGPT

Open “Settings” from your account name at the bottom left
Open “Data controls”
Turn off “Improve the model for everyone”

Gemini

Turning off “Keep Activity” under “Activity” in “Settings & help” stops your chats from being used for training, but your past chats also disappear from the list. So I leave Activity on and use a temporary chat only when I’m working with someone else’s images.

Click “Temporary chat” (the pen icon in a dotted circle) at the top right
Use the input box as usual

Temporary chats aren’t used for training and don’t appear in your chat list. You can’t look at them later, so save any prompts that worked somewhere else.

Grok

Grok and X have separate settings. Turning it off in one doesn’t carry over to the other, so set it in whichever you use.

grok.com: Open “Settings” → “Data Controls” and turn off “Improve the Model”
X: Open “Privacy and safety” in Settings, open “Grok & Third-party Collaborators” under “Data sharing and personalization”, and uncheck the box there

PixAI

PixAI’s privacy policy says it “does not use user-uploaded images to train or improve its foundational AI models without explicit user consent” (as of October 2026).

ComfyUI

ComfyUI generates images on your own PC, so the images and prompts you give it aren’t sent anywhere unless you use API nodes that generate on an outside service. No settings needed.

Character design

If you’re making your own avatar from nothing, you start by thinking about what kind of character it should be. You tell the AI what you’ve decided, such as gender, motif, age, hairstyle, eye color, and build, let the AI fill in the rest, and bounce ideas back and forth until you get a character you like. Here I roughly compare which AI makes the best partner for that.

This time I do have an image to start from, but I’ll use it in the next part, “Regenerating from an image”. Here I assume you’re listing the features of a character that’s only in your head. Since this is character design, I don’t specify a background and leave the pose to the AI.

I list these features: blonde hair, half-up twin tails, golden eyes, maro eyebrows (short, round eyebrows), a mole to the right of the mouth, teen, military-style dress, and a rifle on her back.

ChatGPT

Prompt

I’m working on the design of an avatar for streaming. Please output an image of a character with these features: blonde hair, half-up twin tails, golden eyes, maro eyebrows, a mole to the right of the mouth, teen, military-style dress, and a rifle on her back.

I put the mole in the instructions as an experiment, but it’s faster to add a mole yourself than to have the AI output it. It’s just one dab of a dark skin color.

A character design sheet. At the top left, a list of the requested features and a full-body drawing; in the middle, a close-up of the upper body; to its right, the back view; on the far right, five expressions; along the bottom, close-ups labeled eyes (gold), maro eyebrows, mole (right of mouth), half-up twin tails, and rifle on the back, plus a color palette. Blonde hair tied with black ribbons, half-up twin tails in the front views and twin tails in the back view. A black military-style dress with black thigh-high socks and lace-up boots — As instructed

A character design sheet. A close-up of the face at the top left, a full-body drawing in the middle, six expressions at the top right, below them close-ups labeled eye close-up, mouth (mole on the right), maro eyebrows, back (with rifle), uniform details, and rifle (details) with a list of features, and a small back view and color palette at the bottom left. A black beret, blonde hair tied with black ribbons, half-up twin tails in the front views and twin tails in the back view. A black military-style dress with long black boots — With “baby face” added

Maybe half-up twin tails is a hairstyle it can’t do reliably: you can just see some hair left down at the back in places, but at a glance they’re basically twin tails. It also seems better to specify the eyebrow color. The mole comes and goes and multiplies, so it isn’t stable.

The face looked grown-up, so I added “baby face” to the features and output another one, and it came out cute. You can ask for a baby face, but if you give an age, it tells you off: “We’re very sorry. Your prompt may violate our content policy. If you think this is a mistake, please try again or edit your prompt.” ChatGPT tends to say no to all sorts of things once children are involved. Still, it outputs cute images, so as long as you’re not making a character that shows a lot of skin, it passes.

ChatGPT hits its daily image generation limit fairly quickly, so if you want to generate lots of images, paying will make things more comfortable. I can just wait until the next day, so I personally don’t feel much need to pay.

Gemini

I output it with the same instructions as ChatGPT.

A wide image. In the front yard of a stone building in the evening, the same character appears in full body and upper body. Wavy blonde twin tails, thick eyebrows, and a mole by the mouth. A navy and gold military-style dress with a white frilled hem, black socks and lace-up boots, a rifle with a wooden stock on her back, laughing with a hand on her hip — As instructed

A wide image. In a child's room with bookshelves and stuffed toys, the same character appears twice in full body and once in upper body. Blonde twin tails with an ahoge and hairpins, thick eyebrows, a mole by the mouth, smiling. A navy and gold military-style dress with a red ribbon, a cream frilled hem, black thigh-high socks and brown boots, and a rifle with glowing blue parts on her back — With “baby face” added

Almost the same composition and outfit as the image on the left, with only the face changed. Thinner, lighter eyebrows and a young face with the mouth slightly open — With “about 7 years old” added

Not quite as cute as ChatGPT’s, is it… But this one doesn’t refuse when you give an age, so after the baby face instruction I added “make her about 7 years old”, and it output a much younger face. The mole stays in the same place too. I asked for the right of the mouth, and it seems to read that as the right side from the viewer’s point of view. And half-up twin tails don’t seem to get through here either. They’re all twin tails.

I felt it could do better, so I had ChatGPT write a prompt for Gemini.

Prompt

Please draw an original character in an anime and manga style.

Hairstyle and hair color: Blonde. Half-up twin tails, with the hair on each side tied. Soft, fluffy hair.
Eyes: Bright, shining golden eyes.
Eyebrows: Maro eyebrows. Short, separate eyebrows near the inner ends, with no tails.
Facial features: A baby face that still looks young. A small mole on the right side of the mouth.
Age and feel: A teenage girl. Small and cute.
Outfit: A dress based on a military uniform. Mainly black or dark gray, with gold trim, buttons, and a belt. Balance the sharpness of a military uniform with a girl’s cuteness.
Weapon: A rifle on her back.

Make it an appealing character design with a personality that meets these features. Feel free to come up with the details, colors, hair accessories, and the silhouette of the outfit so that everything fits together naturally.

I won’t specify the pose, expression, composition, background, or art style. Keep the character’s features and finish it as a single illustration with your own ideas.

A wide image. In a square in front of a palace-like building at dusk, the same character appears in full body and upper body, with students in uniform in the background. Fluffy pale blonde twin tails with black ribbons, a mustard beret, thick eyebrows, a mole by the mouth, laughing with her mouth open. A black and gold short-sleeved military-style dress with a red tie, black gloves and boots, a saber at her waist. A rifle with 073 ACADEMY written on the wooden stock over her shoulder — Output from the prompt ChatGPT wrote

Not much different. If you like this art style, this might be fine…

Gemini also outputs JPG, so compared with PNG, the rough image quality when you zoom in bothers me too.

Grok

I output it with the same instructions as ChatGPT. Truth be told, I’ve hardly ever used Grok. It was in the news for the wrong reasons for a while, so I just kind of avoided it.

On a white background, a full-body view from behind, looking back. Long, loosely wavy blonde twin tails with dark green ribbons and golden eyes. A dark green and gold military-style dress with gold epaulets and a black frilled hem, a black assault rifle on her back, black thigh-high socks and heeled lace-up boots — Grok on X

Standing from the thighs up in front of a blurred base wall with the words GOLDEN EAGLE STRATEGIC COMMAND and an eagle emblem. Blonde twin tails with dark green ribbons, smiling with golden eyes, both hands in her pockets. An olive military-style dress with a harness and pouches, an assault rifle hanging from her shoulder — Grok on X with “baby face” added

On a white background, facing forward from the thighs up. Shoulder-length blonde twin tails with navy ribbons, golden eyes, a mole by the mouth. A navy and gold military-style dress with gold aiguillettes and a star medal, a brown leather belt and a cross-body strap, and a rifle with a wooden stock hanging from her right shoulder — grok.com with “baby face” added

Even with the same instructions, the feel of the outfit changed every time. Half-up twin tails don’t come through here either. It doesn’t seem to understand maro eyebrows either. Those are just thick eyebrows. The baby face instruction barely seems to do anything. The art style is close to ChatGPT’s, I’d say. I’d heard Grok is better at photorealistic images, so I think it output cuter images than I expected.

As with Gemini, I had ChatGPT write a prompt for Grok. I run it on grok.com.

Prompt

Please draw an original anime-style character.

Character features

Blonde hair.
Half-up twin tails. The hair near each side of the head is tied, and the rest hangs down naturally.
Bright, shining golden eyes.
Maro eyebrows. Short, separate eyebrows near the inner ends, with no tails.
A small mole on the right side of the mouth.
A teenager. A baby face with a cute, young look.
A dress based on a military uniform. Mainly black or dark gray, with gold buttons, trim, and a belt. A design that combines the sharpness of a military uniform with a girl’s cuteness.
A rifle on her back.

Design direction

Be sure to include all the features above and finish it as a unified, appealing character. Feel free to design the hair accessories, outfit details, decorations, and color balance.

I won’t specify the pose, expression, composition, background, camera angle, or art style. Keep the character’s features and generate a single, polished illustration with your own ideas.

Draw it as a regular character illustration, not a character reference sheet or design sheet.

The Imagine screen on grok.com. The entered prompt at the top, and two tall images below it. Both show a smiling girl with long blonde twin tails, black ribbons, and golden eyes, in a black and gold military-style dress with a white frilled hem and a rifle with a wooden stock on her back. On the left she faces forward with a cape on her shoulders; on the right she looks back with a flag in the background — Two images output from this prompt

That’s a sudden jump in quality. Part of it is probably that I chose quality over speed when generating. One of them actually has half-up twin tails.

But at this point it showed an image generation limit error, and I couldn’t generate any more. If I hadn’t chosen quality, I might have gotten one more try.

PixAI

PixAI uses points to generate, so if you want to enjoy it for free, collect the daily login bonus and save up points. Just looking at other people’s images on the top page is fun, so I find myself opening it at least once a day without thinking. Some people share the prompts and LoRAs they used, so you can also see how an art style you like was made.

From here on I can’t use the same prompts, so I prepare prompts written for image generation. The output probably varies a lot depending on the model, so this may not be a perfect comparison.

Positive prompt

masterpiece, best quality, anime style, 1girl, solo, original character, teenage girl, youthful face, baby face, blonde hair, half-up hairstyle, two side ponytails, hair tied on both sides of the head, long loose hair, hair flowing down over shoulders, golden eyes, mayadare eyebrows, short eyebrows, mole near the right corner of the mouth, military uniform dress, black military dress, gold trim, gold buttons, military decorations, belt, pleated skirt, rifle strapped to her back, detailed character design, beautiful detailed eyes, intricate outfit details

Negative prompt

worst quality, bad quality, low quality, lowres, blurry, bad anatomy, malformed hands, extra fingers, missing fingers, extra limbs, multiple girls, multiple people, text, watermark, signature, adult woman, old woman, rifle in hands, full twin tails, high ponytail, single ponytail, hair tie in mouth, hair elastic in mouth, holding hair tie with mouth, object in mouth, biting object, tying hair

Since I’ll be passing a reference image later, I use the Haruka v2 model. The model PixAI is pushing right now is the Tsubaki series, but Tsubaki has no field for a reference image. I don’t use any LoRA, pass only the positive and negative prompts, and leave everything else at the defaults.

Standing from the thighs up in front of blue lockers. Long blonde hair with the top part tied on both sides with black ribbons in half-up twin tails, short eyebrows, golden eyes, a mole under the eye. A black and gold military-style dress with a brown leather belt and a cross-body strap, and a rifle with a wooden stock on her back

Standing from the thighs up against a white background framed in gold. Long blonde hair with the top part tied on both sides with black ribbons in half-up twin tails, short eyebrows, golden eyes, a mole under the eye. A black and gold military-style dress with a brown leather belt, and the rifle on her back showing

From the thighs up against a blue sky inside a gold frame. Long blonde hair with the top part tied on both sides in half-up twin tails, short eyebrows, golden eyes, moles under the eye and on the cheek. A black and gold military-style dress with gold aiguillettes, and a rifle with a wooden stock on her back

Standing from the thighs up in front of a metal wall. A black military cap, long blonde hair with the top part tied on both sides in half-up twin tails, short eyebrows, golden eyes, a mole under the eye. A black and gold military-style dress with a brown leather belt and a cross-body strap, and a rifle with a wooden stock on her back

The mole position isn’t under control, but the half-up twin tails and maro eyebrows, which tend to get ignored, come out consistently. Then again, I rewrote the prompt several times to get them, so I can’t really call it consistent… So, just to see, I’ll try a prompt in Japanese.

With “half twin tails” in Japanese it seemed to think I meant pigtails and added braids, so I changed it to “two side up”. Writing “baby face” made her hold a baby, so I dropped baby face and teen and used “child” instead.

Japanese positive prompt

子ども、金髪、ツーサイドアップ、金色の目、麻呂眉、口元右にほくろ、軍服ワンピース、背中にライフル

In English: child, blonde hair, two side up, golden eyes, maro eyebrows, a mole to the right of the mouth, military-style dress, a rifle on her back. No negative prompt.

Facing forward from the chest up in front of a window with shrubs outside. Long blonde hair with the top part tied on both sides with black ribbons in half-up twin tails, short eyebrows, golden eyes, a sulky face, a mole by the mouth. An olive military-style dress with a cross-body strap, and a rifle with a wooden stock on her back

Standing from the thighs up in front of a wall with stone pillars and flowerpots. A blue military cap, long blonde hair with the top part tied on both sides in half-up twin tails, short eyebrows, golden eyes, a sulky face, a mole by the mouth. A blue and white military-style dress with a cross-body strap, and a rifle with a wooden stock on her back

Facing forward from the thighs up with a white stone wall behind her. Long blonde hair with the top part tied on both sides with black ribbons in half-up twin tails, short eyebrows, golden eyes, a sulky face, a mole by the mouth. A white and gold military-style dress with a brown cross-body strap and a pouch, and a rifle with a wooden stock on her back

Facing forward from the chest up on a cobblestone path. Long blonde hair with the top part tied on both sides with black ribbons in half-up twin tails, short eyebrows, golden eyes, a mole by the mouth. A dark green and gold military-style dress with a cross-body strap, and a rifle with a wooden stock on her back

There it is… I should’ve just used Japanese from the start… Come to think of it, when you open the generator, the input box has a Japanese placeholder example like “a girl reading a book under a cherry tree”. PixAI has a Japanese subsidiary in Tokyo, so the Japanese pages seem to be run from Japan. It looks like Japanese prompts get converted automatically.

At this point it hit me: “Maybe ‘two side up’ gets through better than ‘half-up twin tails’?” So I output again with ChatGPT, Gemini, and Grok, but it didn’t make much difference. ChatGPT gave almost the same image as before, Gemini still gave twin tails, and Grok gave half-up buns.

ComfyUI

I’ve written up how to install ComfyUI here.

Setting up AI image generation on your own PC Getting ComfyUI running on Windows: two ways to install it, picking a model and checking its license, and how the nodes fit together. GyunGyun×Create (gyungyun.com)

Japanese prompts obviously won’t work here, so I start over from the prompt. The model is WAI-illustrious-SDXL v16. Apparently v17 is out, but I haven’t updated, so I’ll stay on v16 for this note. After some fine-tuning, I ended up with this.

Positive prompt

masterpiece, best quality, amazing quality, sfw, 1girl, solo, child, cute, smile, blush, :3, looking at viewer, blonde hair, medium hair, two side up, black hair ribbon, golden eyes, short eyebrows, mole under mouth, military uniform, military dress, black dress, gold trim, gold buttons, belt, pleated skirt, rifle on back, sling

Negative prompt

nsfw, large breasts, lowres, bad quality, worst quality, worst detail, sketch, censor, bad anatomy, bad hands, extra fingers, missing fingers, multiple girls, multiple views, text, watermark, signature, twintails, long hair, very long hair, holding gun, angry, frown, serious, v-shaped eyebrows, glaring, sharp eyes

Facing forward from the thighs up in front of a dark door. Shoulder-length blonde two side up with black ribbons, short eyebrows, golden eyes, smiling with a :3 mouth, a mole by the mouth. A black and gold military-style dress with red accents, a brown cross-body strap, and a rifle with a wooden stock on her back

Facing forward from the thighs up in front of a gray door. Shoulder-length blonde two side up with black ribbons, short eyebrows, golden eyes, smiling with a :3 mouth, a mole by the mouth. A black and gold military-style dress with gold aiguillettes and a red armband, a brown cross-body strap, and a rifle with a wooden stock on her back

Standing from the thighs up in a bright hallway. Shoulder-length blonde two side up with black ribbons, short eyebrows, golden eyes, smiling with a :3 mouth, a mole by the mouth. A black and gold military-style dress with medals, a brown cross-body strap, and a rifle with a wooden stock on her back

Standing from the thighs up in front of a wooden wall. Shoulder-length blonde two side up with black ribbons, short eyebrows, golden eyes, smiling, a mole by the mouth. A black and gold military-style dress with a red armband, a brown cross-body strap, and a rifle with a wooden stock on her back

I put sfw in so it doesn’t turn into an adult image. Without a hair length, half-up twin tails come out long, so I specified medium hair. Getting too specific would stray from the point, so I’ll stop here for now.

I think PixAI’s Japanese prompt would also have given something similar if I’d specified hair length and expression.

Character design summary

AI	What was good	What bothered me
ChatGPT	Outputs cute images even from loose instructions	Half-up twin tails and the mole aren’t stable. Refuses under its content policy if you give an age. Hits the daily image generation limit quickly
Gemini	Doesn’t refuse when you give an age. Keeps the mole in place	Not as cute as ChatGPT. Half-up twin tails don’t get through. Not much different even with a prompt written by ChatGPT. JPG, so the image is rough when you zoom in
Grok	With a prompt written by ChatGPT, run on grok.com with quality chosen, the quality goes up and half-up twin tails appeared	The feel of the outfit changes every time, even with the same instructions. Maro eyebrows and baby face don’t work. Hits the daily image generation limit quickly
PixAI	A Japanese prompt gave almost exactly what I asked for	With English prompts, it took several rewrites to get half-up twin tails and maro eyebrows. The mole position can’t be controlled. Free, but uses points
ComfyUI	Half-up twin tails came out once I specified medium hair	Japanese prompts don’t work, so you have to fine-tune English prompts