So you want to create this gorgeous elf girl holding a bow — ethereal, glowy, Danbooru-tag perfection —
✨ NOT whatever strange cursed artifact appears when you type “beautiful elf” and pray for the best. ✨

Welcome.
This is your oversized, too-long, wiki page for diving into AI art, NSFW workflows, Stable Diffusion, SDXL, Pony Diffusion, LoRAs, samplers, VAEs, and all the little gremlins inside the model that determine whether your waifu looks majestic or… melted.

Everything here is explained in plain English, with minimal techno-elitist nonsense.
And yes — this is the same tech behind uncensored, NSFW-ready models like the ones on Nemora.ai.

Let’s get started.

🌈 Base Models Explained (SD 1.5, SDXL, Pony, Flux)

🟦 Stable Diffusion 1.5 (SD 1.5)

The nostalgic classic. The fandom darling.
If SDXL is a DSLR, SD 1.5 is the slightly scuffed smartphone camera that still slays.

Why people love it:

• Runs on potato hardware (4–6GB VRAM)
• The widest NSFW and anime ecosystem
• LoRAs everywhere
• Slightly stylized → perfect for fandom art

Use for: NSFW, anime, stylized characters, quick prompting.

🟩 Stable Diffusion XL (SDXL)

The ✨high-definition bitch✨.

Why it’s better:

• Sharper anatomy
• More realism
• Gorgeous lighting
• Better prompt comprehension

BUT!!
Requires 12–16GB VRAM.
If your GPU is a wild animal you found behind a dumpster → not SDXL-friendly.

🟧 Pony Diffusion (SDXL-Based)

No, it’s not just ponies.
It’s the emotionally expressive anime girl factory.

Great for:

• Anthro, furry, stylized anime
• Character consistency
• Dramatic expressions
• Niche fandom aesthetics

Works beautifully for NSFW and very specific styles.

🟥 Flux (1.dev & 1.S)

Newest. Smartest. Also… not recommended for spicy art right now.

Why?

• ✘ Restrictive licensing
• ✘ Distilled version can’t be fine-tuned properly
• ✘ Limited NSFW checkpoint ecosystem
• ✘ Big model → big GPU needed

Flux is strong, but not fandom-friendly just yet.

TL;DR — Which Base Model to Pick?

• SD 1.5 → Maximum NSFW compatibility
• SDXL → Best realism
• Pony → Best anime/furry/expressive characters
• Flux → Not ideal for NSFW yet

🧬 Checkpoints, Fine-Tunes, Merges & VAEs — The Customization Playground

Checkpoint = The “Main Personality” of the Model

Examples: Anything V5, Realistic Vision, Pony Diffusion.

This is the file you load into your UI.
Think of it as picking a playable class before you start generating.

Fine-Tune = Specialization

Fine-tunes retrain a base model to be better at something specific:

• NSFW
• Anime girls
• Thick thighs™
• Elves
• Latex photography
• Literally anything

Most good NSFW models = fine-tunes.

Merge = Frankenmodel

Two or more checkpoints blended together.

This can produce:

• ❗ Absolute magic
  or
• ❗ Absolute chaos

Great for style-hunters.

VAE = The Color/Contrast/Sharpness Final Boss

VAEs fix dull images, washed-out colors, blurry skin, gray undertones.

If your characters look… ashy?
It’s the VAE. Always the VAE.

Common VAEs:

• vae-ft-mse-840000 → safe SD1.5 default
• Custom Pony/Anything/AOM VAEs → recommended by the creators

🦋 LoRAs, QLoRAs, DORAs & Embeddings

These are ADD-ONS you load on top of your checkpoint to give it new skills.

💛 LoRA — The Supreme Overlord

The MOST important tool in AI prompting.

LoRAs customize your model for:

• Characters (OC, waifu, celeb)
• Art styles
• Body types
• Kinks/NSFW details
• Specific poses or clothing behaviors

You activate one like:
<lora:character:0.8>

And bam — new skill unlocked.

You can stack LoRAs.
Yes, make Spider-Gwen adjusting her clothes with a Dreadlocks LoRA.
Dream big.

💚 QLoRA

For making LoRAs efficiently.
You DON’T need it unless you’re training your own.

💜 DORA

LoRA but dynamic, more complex, more experimental.

Mostly ComfyUI territory.

🩵 Embeddings (Textual Inversions)

Tiny keyword modifiers for:

• Textures
• Lighting
• Very specific faces

They’re subtle. Like seasoning, not sauce.

TL;DR

• LoRA → You NEED this.
• QLoRA → Only for training.
• DORA → Power user toy.
• Embeddings → Small tweaks.

🎨 Interfaces — A1111, Forge & ComfyUI

AUTOMATIC1111 (A1111)

The OG. The fan favorite.

Pros:
✔ Simple
✔ Tons of NSFW extensions
✔ Perfect for beginners

If you’re new → start here.

WebUI Forge

A cleaner, faster A1111.

✔ Faster model loading
✔ Better SDXL performance
✔ Same extensions

Great upgrade.

ComfyUI

The “visual programming” playground of image generation.

✔ Node-based workflows
✔ Super powerful
✔ Great for advanced setups
✘ A bit much for beginners

TL;DR

• New? NSFW? → A1111
• Want faster SDXL? → Forge
• Want full pipeline control? → ComfyUI

📝 Prompts 101 — Positive vs Negative

Positive Prompt = What you want

Describe:

• Subject
• Style
• Lighting
• Pose
• Mood
• Camera angle

Example:
blonde anime girl, red eyes, neon alley, leather outfit, volumetric lighting, high detail

Negative Prompt = What you DON’T want

Keep basic negatives handy:

blurry, extra fingers, bad anatomy, watermark, mutated hands

Add more when problems appear.

🗂️ How to Write Good Prompts (The Tumblr Way)

Bad prompt:

“beautiful elf warrior”

The model shrugs. Elf where? Doing what? Wearing what?

Better prompt:
elf warrior, silver armor, forest, glowing eyes, volumetric fog, holding sword, full body, masterpiece

This tells the AI:

• Who she is
• Where she is
• What she’s doing
• What the mood is
• What to prioritize

🏷️ Danbooru Tags — The Secret Anime Sauce

Anime models think in tags, not sentences.
Use tags → accuracy skyrockets.

Examples:
1girl, blonde hair, red eyes, armor, bow, forest, dynamic lighting, full body

Avoid full natural language unless you like surprises.

💥 Keyword Emphasis

Use parentheses to emphasize or de-emphasize:

(red eyes:1.3) → more attention
(background:0.4) → less

Stay between 0.1–1.5 unless you enjoy cursed imagery.

⚙️ Samplers, Steps & Why Your Image Looks Weird

Samplers = How The Model Draws

• Euler a → Fast, great for anime
• DPM++ SDE Karras → Most detailed, slower
• DPM++ 2M → Clean realism
• DDIM → Fast & loose

Check your checkpoint page for recommended samplers.

Steps

General sweet spot = 20–40

More steps = fewer mistakes
Too many steps = overcooked, mushy, or noisy

🖼️ Resolution — Yes It Actually Matters

SD 1.5

Trained at 512×512
Best sizes:

• 512×768
• 512×640
• 768×512

Too high → blur city.

SDXL / Pony

Trained around 1024×1024
Sweet spot:

• 1024×1024
• 832×1216
• 1152×896

🧰 Plugins to Level Up Your Work

ADetailer

Auto-fixes:

• Faces
• Hands
• Eyes

Essential for NSFW creators.

ControlNet

Lets you control:

• Pose
• Depth
• Lineart
• Composition

Perfect for consistent characters or reference-based art.

🚑 Troubleshooting Cheat Sheet

• Gray/washed-out image → Wrong VAE
• Weird hands → More steps + ADetailer
• Model ignoring prompt → Use tags
• Flat lighting → Add lighting descriptors
• Character inconsistency → Add LoRA

🌙 Wrapping Up (For Now…)

Congrats — you now understand:

• Base models
• Checkpoints vs fine-tunes
• LoRAs vs embeddings
• Prompt structure
• Samplers
• Resolutions
• Why Flux is a diva
• Why VAEs are secretly the final boss

AI image generation is a playground — and you now have the keys.