Techalicious Academy / 2025-12-11-stable-diffusion

(Visit our meetup for more great tutorials)

SETTINGS EXPLAINED

When you generate an image, you're not just typing a prompt. There's a whole control panel of settings that affect how the AI works. Here's what each one does in plain English.

SAMPLER METHOD

What it is:

The mathematical recipe the AI uses to build your image from noise.
Think of it like different brush techniques - same paint, different
results.

The ones we use:

euler_karras
  Smooth, consistent results. Good all-rounder. We use this for
  Bridgetoons and Arthemy.

dpm++_karras
  Slightly sharper details, good for photorealistic stuff. We use
  this for Cyberrealistic and Realism.

euler_a (euler ancestral)
  More random/creative variation between generations. Same seed
  won't give identical results. Good for exploration.

The "_karras" part:

A noise schedule that improves quality. If you see a sampler with
and without "_karras", pick the karras version.

Bottom line:

Just use what's recommended for each model. These were tested.

STEPS

What it is:

How many times the AI refines the image. Starts as pure noise,
gets clearer each step.

Too few (under 20):

Image looks rough, unfinished, maybe blurry.

Sweet spot (25-45):

Where most good images happen. More steps = more detail, but
diminishing returns after a point.

Too many (over 60):

Wastes time. Image might get "overcooked" - weird artifacts.

Our models:

Realism: 27 steps
Cyberrealistic: 35 steps
Bridgetoons: 38 steps
Arthemy: 45 steps

Why different numbers? Each model was trained differently and hits its sweet spot at different step counts.

CFG SCALE (Classifier-Free Guidance)

What it is:

How strictly the AI follows your prompt versus doing its own thing.

Low CFG (3-5):

AI takes more creative liberty. Softer, more natural looking.
Good for photorealistic models.

Medium CFG (6-8):

Balanced. Follows prompt well but still looks natural.

High CFG (9-15):

Very literal interpretation. Can look artificial or "crunchy"
if pushed too high. Colors might oversaturate.

Our models:

Realism: 4.0
Cyberrealistic: 4.5
Bridgetoons: 6.0
Arthemy: 6.0

Notice the photorealistic models use lower CFG - they need that creative freedom to look natural.

CLIP SKIP

What it is:

CLIP is the part that translates your text into something the AI
understands. "Clip skip" means skipping some layers of that process.

Clip skip 1:

Uses all layers. More literal, sometimes too literal.

Clip skip 2:

Skips the last layer. More artistic interpretation. This is what
anime and illustration models prefer.

Our models:

ALL USE CLIP SKIP 2

Set it once and forget it for these models.

SEED

What it is:

The random starting point for generation. Same seed + same settings
= same image (mostly).

How to use it:

-1 = random seed each time (default, good for exploring)
Specific number = reproduce that exact result

Pro tip:

Find an image you like? Note the seed. Then tweak the prompt
slightly and keep the seed to iterate on that composition.

RESOLUTION (Width x Height)

What it is:

Image dimensions in pixels.

Why it matters:

Models are trained on specific resolutions. Stray too far and
you get weird results - stretched faces, duplicated body parts.

Our models:

Bridgetoons: 672x1200 (portrait)
Cyberrealistic: 832x1216 (portrait)
Realism: 896x1152 (portrait)
Arthemy: 1024x1344 (portrait)

Stick to these. If you need landscape, swap width and height.

HIRES FIX (High Resolution Fix)

What it is:

A two-pass process. First generates a smaller image, then upscales
and adds detail in a second pass.

Why it exists:

SDXL struggles with fine details (especially eyes and hands) when
generating large images directly. Hires fix works around this.

The settings:

Hires upscale: 2.0
  How much bigger. 2x means a 672x1200 image becomes 1344x2400.

Hires steps: 15
  How many refinement steps in the second pass.

Hires upscaler: Lanczos
  The algorithm used to enlarge. Lanczos is sharp and clean.

Denoising strength: 0.2-0.3
  How much the second pass can change things. Lower = closer to
  original. Higher = more freedom to add/change details.
  - 0.2 = subtle refinement (Arthemy)
  - 0.3 = moderate refinement (Bridgetoons)

Our models with Hires Fix:

Bridgetoons: YES (2x, Lanczos, 0.3 denoise, 15 steps)
Arthemy: YES (2x, Lanczos, 0.2 denoise, 15 steps)

Our models without Hires Fix:

Cyberrealistic: NO
Realism: NO

Why some models don't use it:

On Apple Silicon (Macs with M1/M2/M3), hires fix can produce
artifacts - blurry patches or weird scan lines. We disable it
for models where it caused problems.

BATCH SIZE AND BATCH COUNT

Batch size:

How many images to generate simultaneously. Uses more memory.
Keep at 1 unless you have tons of RAM/VRAM.

Batch count:

How many times to run generation. Makes images sequentially.
Set to 4 to get 4 different variations of your prompt.

PUTTING IT ALL TOGETHER

When you're starting out:

  1. Pick your model
  2. Use its recommended settings (sampler, steps, CFG, resolution)
  3. Set Clip Skip to 2
  4. Leave seed at -1 for random
  5. Enable Hires Fix only for Bridgetoons and Arthemy
  6. Write your prompt and hit Generate

Once you get comfortable, experiment. But start with the known-good settings so you have a baseline.