Michael Limberger

Need me? Email mike@limberger.ca

The rest of the panel

When you generate an image, you are not just typing a prompt. There is a control panel of settings that change how the AI works. Here is each one in plain English.

Sampler

The sampler is the mathematical recipe the AI uses to build your image from noise. Think of different brush techniques: same paint, different results.

euler_karras is smooth and consistent. Good all-rounder. We use this for Bridgetoons and Arthemy.

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

euler_a (euler ancestral) is more random between generations. Same seed will not give identical results. Good for exploration.

The _karras part is a noise schedule that improves quality. If you see a sampler with and without it, pick the karras version. Use what was tested for each model.

Steps

Steps are how many times the AI refines the image. It starts as pure noise and gets clearer each step.

Under 20 looks rough, unfinished, maybe blurry. The sweet spot is 25 to 45. More steps mean more detail, with diminishing returns. Over 60 wastes time. The image might get overcooked: weird artifacts.

Realism: 27 steps. Cyberrealistic: 35. Bridgetoons: 38. Arthemy: 45. Each model was trained differently and hits its sweet spot at a different count.

CFG scale

CFG is Classifier-Free Guidance: how strictly the AI follows your prompt versus doing its own thing.

Low CFG (3 to 5): more creative liberty. Softer, more natural. Good for photorealistic models. Medium CFG (6 to 8): balanced. Follows the prompt well but still looks natural. High CFG (9 to 15): very literal. Can look artificial or crunchy. Colors might oversaturate.

Realism: 4.0. Cyberrealistic: 4.5. Bridgetoons: 6.0. Arthemy: 6.0. The photorealistic models use lower CFG. They need that freedom to look natural.

Clip Skip

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. This is what anime and illustration models prefer.

Our models all use clip skip 2. Set it once and forget it.

Seed

The seed is the random starting point. Same seed plus same settings gives the same image, mostly.

-1 is a random seed each time. That is the default, and it is good for exploring. A specific number reproduces that exact result.

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

Resolution

Width by height in pixels. Models are trained on specific resolutions. Stray too far and you get weird results: stretched faces, duplicated body parts.

Bridgetoons: 672x1200 portrait. Cyberrealistic: 832x1216. Realism: 896x1152. Arthemy: 1024x1344. Stick to these. If you need landscape, swap width and height.

Hires Fix

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

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

Hires upscale 2.0 means a 672x1200 image becomes 1344x2400. Hires steps 15 is how many refinement steps in the second pass. Upscaler Lanczos is sharp and clean.

Denoising strength 0.2 to 0.3 is how much the second pass can change things. Lower stays closer to the original. 0.2 is subtle (Arthemy). 0.3 is moderate (Bridgetoons).

Bridgetoons: yes, 2x, Lanczos, 0.3 denoise, 15 steps. Arthemy: yes, 2x, Lanczos, 0.2 denoise, 15 steps. Cyberrealistic: no. Realism: no.

On Apple Silicon (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 is how many images to generate at once. Uses more memory. Keep at 1 unless you have tons of RAM or VRAM.

Batch count is how many times to run generation, sequentially. Set to 4 to get 4 different variations of your prompt.

Start here

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

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