Relly's Really Short Image Making Guide for PixAI

Hello, Relic here! Getting started on Janitor is tough, getting an image you like, even more so. But, luckily with some help from friends and hours of tinkering I've made a somewhat solid guide for starting with PixAi. I hope you find this useful, and if you have any questions feel free to join my co-owned Discord Server
PFP

Relictus.

Terminology / Key Words

You don’t need to know all of this if you’re following the guide as-is, but if you want to customize your workflow, this stuff matters. Feel free to skip if you're already familiar.

Prompt:
The text that tells the AI what to generate, subject, style, pose, setting, mood, etc. Be specific and clear.
Prompt

Model:
The base AI model that handles the generation. Different models give different aesthetics.
Model

LoRA:
Short for Low-Rank Adaptation. These are add-ons that help the model replicate specific characters, styles, outfits, or features. You can stack and adjust their influence.
LoRA

Size:
The output resolution in pixels. Impacts aspect ratio and detail. Free users are limited to:
3:15, 7:9, 1:1, and 5:3.
Size

Negative Prompt:
Stuff you don’t want. Helps filter out common AI mistakes like extra limbs, unwanted outfits, or the wrong hair color.
Negative

Sampling Steps:
How many refinement passes the AI makes. More steps = more detail, but also costs more tokens. Most models default to 20–40, but you usually don’t need that much.
Steps

Sampling Method:
The algorithm used to generate the image. Affects sharpness, smoothness, and render style. Common picks:

  • DPM++ 2M Karras (balanced)
  • Euler a (faster, stylized)
    Method

CFG Scale (aka GFC):
Controls how strictly the AI follows your prompt.

  • 2–5 = loose
  • 5–7 = balanced
  • 7+ = strict, but can cause jank or artifacts
    CFG

Seed:
Controls the randomness. Use the same seed with the same settings and you’ll get the same image every time. Leave blank for variety.
Seed

VAE Model:
Stands for Variational Autoencoder. Handles color and fine detail. Most of the time you can just leave it on the model’s default.
VAE


Step 1: Creating a Burner Email

Always use a burner or spare email. PixAI sends way too many marketing emails, and there’s no reason to connect your personal inbox to it.

Gmail and Outlook are fast options for making throwaway accounts. Just know that after a few, Google will start asking for phone verification, which defeats the purpose.

If you want to avoid all that, use Temp-Mail. It gives you a quick, disposable address you can use to sign up and verify.

⚠️ Warning:
Temp-mail addresses are temporary and unrecoverable. If you lose access to your PixAI account, it’s gone. Use only for accounts you don’t care about long-term.


Step 2: Setting Up Your PixAI Account

  1. Go to https://pixai.art
  2. Sign up with your burner/spare email
  3. Click your profile picture (top right), then choose “Profile”
  4. Click “Edit Profile”
  5. On the left sidebar, go to “Accounts” and verify your email
  6. Then go to “Content Settings” and:
  • ✅ Enable 18+ confirmation
  • ✅ Allow sensitive/flagged content
  • ❌ Disable SafeSearch
  • ❌ Turn off blur sensitive content

This is important if you plan to generate NSFW or borderline images.


Step 3: Writing a Good Prompt

This is where most people mess up. PixAI is dumb. It doesn’t read your prompt like a human. It doesn’t "understand" story beats or poetic language. It just matches tags from its training data.

Bad example:

"A young woman in her mid-twenties leaning against a brick wall with her face obscured under the night sky."

Sure, it might generate something decent, but it’s vague. You’re giving the AI too much control, and it will usually get at least one thing wrong.

Better:

1girl, long hair, black hair, messy hair, hidden face, shadows, dark lighting, night, brick wall, solo, street

Short, stacked tags. Each one tells the AI to focus on a specific element. The more control you take, the better the outcome.


Pro Tips


Useful Tag Types

  • 1girl, solo, smile, looking_at_viewer
  • blonde hair, twin tails, green eyes, school uniform
  • night, sunset, cyberpunk, indoors, urban
  • cinematic lighting, dark shadows, moody, bloom

Try to stick between 8–15 focused tags. Then adjust from there.


Step 4: Generation Settings

Here’s what I use most of the time for 3:5 ratio, 4-image batches with high priority:

  • Sampling Steps: 16–20 (costs 2,200–2,800 tokens)
  • Sampling Method: Euler a
  • CFG Scale: 5–7
  • VAE: Default

A lot of models default to higher settings than they need. These strike a good balance between quality and cost.


Step 5: Upscaling with Enhance

Once you’ve got an image you like, you can clean it up or sharpen the details using PixAI’s Enhance tool.

How to use it:

  1. Hover over the image
  2. Click the icon in the bottom-left corner (sparkle/magic wand)
  3. Select Image Enhance if it isn't already
  4. Adjust the settings using my steps below.
  5. Click Generate

Ratio

Controls the final size and shape of the image.
I use 1.5x - 1.6x tends to crop off the sides on 3:5 images.

Denoising Strength

Controls how closely the upscale sticks to the original.

  • Lower = more faithful
  • Higher = more altered

I usually stay between 0.25 – 0.30, sometimes up to 0.35 if I want more change.

Denoising Steps

Doesn’t really do much.
I just set it to 10 or 16 and move on.

Prompt Strength

You can adjust how strongly the AI focuses on certain tags by wrapping them:

  • (tag) = more emphasis
  • {tag} = less emphasis
  • ((tag)) = even stronger
  • {{tag}} = even weaker

Example:

(dramatic lighting) → makes it more likely
{blurry background} → makes it less likely

This doesn’t remove the tag like a negative would, it just changes how much the AI pays attention to it. If you want something gone, use the Negative field.


🎨 Artist Tags vs. LoRAs What’s the Difference?

🔹 LoRAs (Universal and Flexible)

LoRAs (Low-Rank Adaptations) are external modules that work with most models. They’re trained on specific characters, styles, clothing, or poses. You can load multiple LoRAs at once and adjust how strongly they affect the final image using a strength value slider.

Basically, they’re like add-ons you can mix and fine-tune.


🔸 Artist Tags (Model-Specific)

Artist tags are built-in style triggers that only work with specific models like Illustrious and NoobAI. These models are trained on massive Danbooru datasets that include artwork from specific artists. That training allows the model to mimic an artist’s style just by using their name as a tag.

For example, if you wanted to generate an image using the style of let's say, afrobull. You'd simply type their artist name in the prompt box like so:

  • kinkymation, higashiyama kobeni, looking at viewer
    And get an image like this!
    Kobeni However, not all artist names will trigger, and some only work if you use their alternate tag or alias. You’ll want to test what works best on the model you’re using.

A full list of tags can be found here. https://rentry.co/zokus-artstyle-repo


🧠 TL;DR

  • LoRAs = Loadable, customizable modules that work with most models
  • Artist Tags = Built-in styles from Danbooru artists, only supported on Illustrious/NoobAI
  • Some tags won’t work unless you use an alternate name (check Danbooru aliases)

🎁 Daily Free Tokens

PixAI gives you 3 solid ways to get tokens for free every day.

1. Daily Login Bonus

Just log in once a day and get 10,000 tokens. Easy.

2. Publish 10 Images

After generating some images, hit Publish. A popup will appear:

Publish Button
Publish Popup

To get the reward:

  • Set Visibility to Private
  • You don’t need a title or description
  • Hit Publish

Do this 10 times, then:

  1. Go to your profile
  2. Find Published Artworks under rewards
  3. Claim your 10,000 token bonus

You can do this every 24 hours.

3. Likes = Extra Tokens

Every like on your public posts gives you 100 tokens, up to 2,000 per day (20 likes). This one’s less reliable, but it's a nice bonus if you're posting stuff publicly.

4. Share on Social Media

You can get 3,000 tokens daily just for pretending to share your images on platforms like X (Twitter), Facebook, Reddit, etc.

  • On desktop: click the share icon on any image, select a platform, wait for the share window to load, and close it.
  • On mobile/app: tap an image > tap the share icon > tap the three dots > select copy link. That alone will trigger the reward.

No actual posting required. Just opening the menu or copying the link is enough.


🔍 Bonus Tip: Learn from Other People’s Gens

One of the best ways to level up is by just browsing PixAI itself. Look through other people’s generations especially ones that show their prompts and LoRAs. You’ll pick up on good phrasing, useful tag combos, and niche terms you might not think of on your own.

And don’t skip the LoRA descriptions. Some creators actually include tips on how to use them, including recommended weight settings. Not all do, but when they do, it saves you a lot of trial and error.


Acknowledgements and closing statement

This guide couldn't have happened without contributions from members of the community including but not limited to.

Rime - The guy I co-own my server with and someone I can't begin to thank enough for everything he's done for me.
Duplex - Worked with me throughout the entire process writing this guide from the perspective of a beginner. Close friend of mine as well.
Skif. - Huge inspiration when it came to genning images, this guy's style was a huge factor into me making this, also a really good friend.
.zoku - Good buddy and expert when it comes to generating images. Main fact checker for everything. Plus his repo is goated. Also gave me the tip about the 3k daily token sharing method
Leaferson - Telling me about the parenthesis trick
PistolPony - Gave the ending quote
KornyPony and Viki(LuvBytes) - While they didn't help with this directly, they have been incredibly helpful on my creator journey and I love them both to death.
Aobion - Fact checker and good pal


To conclude...

"Please experiment with the models, loras, and temps. Sometimes you'll have the perfect model but the sampling steps need to be higher for example, so remember there's endless ways to gen something you need."
PistolPony

Edit
Pub: 06 May 2025 08:45 UTC
Edit: 09 May 2025 03:16 UTC
Views: 1032