Category: Character Creation
On our lovely little platform, we offer a feature that allows you to take a character and place it in a chat or studio to generate images. While this is an amazing tool, creating a character that prod
MiocAI는 Character.AI, Candy AI 등 AI 캐릭터 채팅 앱의 대안입니다.
On our lovely little platform, we offer a feature that allows you to take a character and place it in a chat or studio to generate images. While this is an amazing tool, creating a character that produces consistent and reproducible images can be a challenging task. Therefore, our first act in the character creation series will introduce you to the tools we provide and the skills you need to ensure your character’s consistency, and how to fire-test and trial around with it to be sure.
Now, we refer to the character's defined appearance as the appearance prompt Alas it comes in the shape of multiple prompts if so desired, more on that later.
MiocAI has a few hard rules that matter for image prompts:
Don't make real people. Not celebrities, not your neighbour or wife or husband or any other real individual, consenting or not. We cannot verify that your person consents, and as such we ask you to fundamentally avoid it.
Don't make underage/childlike characters. Not for images, not for conversation, not in a SFW context, never. Theres too many variables we cant account for, and as such we cannot and will not allow you to create underage-like content.
Keep your prompts clearly adult, fictional, and within platform boundaries, especially if you plan to publish. We enforce these rules, and you will have your skin removed and your character deleted. Dont risk it.
Some people, not saying you would, but some people would go off and make their character prompt look a little like:
a girl in a garden while its raining, she is called XYZ and she is XY years old and she is smiling
Now, if you dont know any better, this is a completely reasonable prompt. Except, it isnt for your character. See, when creating a character, youre creating it to be versatile. to be reusable, and to produce the same character in different places, expressions, poses, etc.
If you bake your scene, weather, time, expression, pose into the prompt, youre going to have all of those things in all images. Even when the character is supposed to be on a sunny beach for instance.
So, theres no point in putting the Action/Scene prompt into your character's prompt field. If you cram smirking, lying on bed, cinematic lighting, city skyline, holding sword into the appearance prompt, you’re baking those into EVERY. DAMN. IMAGE. that your character will ever generate. Yuck.
Fun fact, when you choose the 4-prompt (outfit) system in character creation, the "extras" option is locked on purpose. Its there to discourage users from considering ever placing any scene/action related details into the character prompt. Like a honeypot!
That’s how you end up with a character who’s always shot from the same angle or always holding the same object.
One of the fastest ways to get consistency is to start from a clean reference image you actually like.
Why? Studio generation tends to be better for design exploration because you can just generate directly from a prompt. Without having to have a character set. So you can experiment with things that give you the best consistency, per say. Once you got a character that generates consistently for you, you can snatch that prompt, and toss it into character creation. And voilia, youve got yourself a bulletproof character that doesn't shape-shift. (unless thats what youre going for)
Uploading a profile picture is mostly cosmetic. MiocAI doesn’t use your avatar upload as a “reference image” during normal image generation. At least not yet. What controls appearance is the appearance/outfit prompt (and any outfit system you’ve set up).
So if your character’s prompt says “short silver hair,” but your avatar has long black hair,the generator will follow the prompt, not the avatar.
Exception (tooling): Some flows let you turn an image into a prompt (image-to-prompt), which is useful. But that’s converting the image into text, not using the image directly as a reference every time. More on this when we actually implement this.
So, try to avoid this:
I've been yapping your head off giving you a pre-talk, but where do you actually apply all of this new-found knowledge?
Well first off, youre going to click "bots" in the menu, followed by "create". Set your name, hit next, write out personality and backstory (our next article will tackle those), hit next, and voilia. The image settings page.
Once you reach there you’ll see:
Pick one and commit early. Switching style later often requires prompt cleanup because:
This splits your appearance prompt into:
Pretty simple, and tt helps prevent your outfit and face details from stepping on each other.
If you put something in the appearance prompt, you’re asking for it to appear in every image unless the action prompt overrides it.
That includes:
So the question isn’t “Would this look cool?” It’s: “Do I want this in 90% of my images?”
Like what the fuck is that. Who you tryna fool gojo. Thats CLEARLY something in your hand. I'd wager, as a matter of fact, that its a microphone. Not to rub it in, but could it be that you have that IN YOUR APPEARANCE PROMPT? Tsk.
You can write prompts in natural language, but MiocAI image models, with the only current exception being Sotar, generally behave better with tag-style prompts:
A good pattern is reliance on keywords and key phrases, seperated by commas. You dont want to write out a paragraph describing what your character looks like, as if talking to a human. That wastes very limited space (the Clip tokenizer is limited to around 45 words total, so your character shouldnt exceed 28 words!), and wont result in great images either. The models we use were deliberately trained to work with prompts that go a little like girl, blue hair, black eyes.
Natural-ish (works sometimes, discouraged):
A tall woman with pale skin and short white hair wearing a black jacket
Tag-style (usually stronger):
1girl, pale skin, short white hair, black jacket, confident expression
The tag version is clearer and tends to create fewer “interpretation detours.”
Here an example with natural language:
Versus with simpler tags:
Not the best example, but you can see how much shorter the latter prompt is, and that makes a big difference to accuracy in the long term!
SDXL based models weight early tokens more strongly. This is a fundamental issue with the model's training, not the tokenizer in this case, but it means that words closer to the start of the prompt are respected more.
Let me put this into perspective. Assume you have a prompt that goes smiling, peace sign hands, standing, eyes closed, blue hair, black eyes, and one that goes blue hair, black eyes, smiling, peace sign hands, standing, eyes closed. It looks like a minor difference. The key identity parts of the prompt are at the start. But to the model, they make a world's difference. Whats at the start is more likely to be seen in the produced image.
So do this:
1girl, 1boy, solo (optional)Don’t start with a shopping list of tiny details. Your character’s identity should win the first half of the prompt.
Long prompts don’t just “add more detail.” They also create contradictions, dilute priorities and push your strongest features into the “ignored” zone.
Layer A (Core Identity): must never change Layer B (Signature Look): usually consistent Layer C (Nice-to-have): optional, removed first when quality drops
Example (not for copying,just structure):
1girl, short black hair, amber eyes, beauty mark under eyeblack turtleneck, long coat, silver earringssubtle eyeliner, patterned scarf, ringsIf images get messy, delete Layer C first.
When prompting for specific body types, it’s easy to fall into the trap of using explicit anatomical tags that inadvertently hijack the composition. Commands that emphasize rear size often act as "focus magnets," signaling the AI to position the camera behind the character to showcase that specific trait. This limits your creative control, making it difficult to achieve front-facing or profile shots because the model interprets the descriptive tag as a mandatory focal point for the lens.
To maintain, for instance, a voluptuous aesthetic without losing control of the camera angle, pivot toward silhouette descriptors rather than explicit focal commands. Terms like "wide hips," "defined waist," or "curvy build" communicate the desired figure through the character's overall outline. These descriptors allow the AI to render a full or thick figure while keeping the camera placement flexible, ensuring your character remains curvy whether they are facing the viewer, standing in profile, or captured in a wide-angle shot.
If you specify footwear, many models treat it as a hint that you want to see it, which increases:
If shoes matter, use them intentionally:
Instead of specifying shoes, control framing via:
These control composition without locking you into “full body every time.”
Feet in the place of hands.. Im not complaining 🤤
This is the #1 reason people get “samey” images.
If you write:
Pick a small set of iconic elements and let the scene do the rest:
Everything else belongs in scene prompts or outfits.
Structured mode cuts your prompt into 4. 3 if we are strictly talking about usable ones. This adds the benefit of later allowing for outfit changes as they arent bound to the main prompt, and can easily be seperated. Very effective. Also gives you more control over your character overall.
A good clothing prompt could be something like black bomber jacket, white t-shirt, dark jeans, or long red dress, gold trim, black gloves. Where clothing type and color are clearly outlined.
You cant. We dont allow you to set extras. Cope.
Different MiocAI image models have different personalities.
A common pattern you’ll notice:
So you choose based on what you’re doing:
Pick the model that:
Pick the model that:
A practical approach:
MiocAI gives you a built-in advantage: test images.
If you change 6 things at once, you won’t know what fixed (or broke) it.
Even a simple note helps:
Cause: Hair tags too late, prompt too long, model freestyling Fix: Move hair tags into the first 25–40% of the prompt; delete weak extras
Cause: Outfit list is too detailed or contradictory Fix: Reduce outfit to core items + palette; move “fashion details” to outfit system/outfits
Cause: Shoes, “full body,” overly detailed clothing, body-part focus tags Fix: Remove footwear; add framing tag like “portrait” or “upper body”
Cause: Angle tags in appearance prompt (“from behind,” “from below,” etc.) Fix: Remove angle tags from appearance prompt; use scene/action prompts instead
Cause: Vague tags that imply props (“idol,” “performer,” “soldier”) Fix: Replace vague role tags with visual traits (“stage outfit” vs “performer”)
Below are example prompt styles designed to show structure. Swap details to fit your character.
1girl, short auburn hair, green eyes, pale skin, freckles, black turtleneck, long grey coat
Why it works:
1boy, messy silver hair, blue eyes, small scar on cheek, slim build, white shirt, dark vest, thin tie, round glasses
This will still work, as the elements added (the accessories and the scar) do not use up significant length, and are positioned by importance to the core identity, with the scar being mentioned before the glasses, as the scare is an inheritent part of the character identity. My neck hurts from writing.
Clothes:
black hoodie, dark cargo pants
Body:
athletic build, broad shoulders, relaxed stance
Face:
short black hair, grey eyes, sharp jawline
Extras (optional):
soft lighting
1girl, long brown hair, hazel eyes, curvy build, wide hips, fitted sweater, high-waisted skirt, hoop earrings
Why it works:
📷 Image placeholder: “Curvy silhouette with varied camera angles.”
Some systems support strengthening/weakening tokens (often {} or []). If your model respects it, it can help when one detail keeps getting ignored.
Use sparingly:
Some models respond to generic “quality” tags. They’re not magic, but can help sometimes:
Tradeoff: they consume prompt space and can be placebo on some models. Test once,keep only if it clearly improves results.