Last Updated: November 21, 2025
Sampling Methods
Euler a
Ancestral, creative, non-deterministic
Euler
Simple, fast, deterministic
DPM++ 2M Karras
High quality, good detail
DPM++ SDE Karras
Very detailed, slower
DDIM
Fast, good for img2img
LMS
Linear multistep, smooth
Heun
Accurate, slow, high quality
PLMS
Pseudo Linear Multistep
UniPC
Fast, unified predictor-corrector
Steps: 20-30 typical
More steps = higher quality, slower
Steps: 50+ for details
Diminishing returns after 50
Karras noise schedule
Better detail preservation
CFG Scale (Guidance)
CFG 5-7: Creative, loose
AI takes more freedom
CFG 7-10: Balanced
Good prompt adherence
CFG 10-15: Strict
Follows prompt closely
CFG 15-20: Very strict
May oversaturate
CFG 20+: Distorted
Often artifacts, avoid
Lower CFG for artistic
More natural, fluid results
Higher CFG for precision
Exact prompt matching
Negative Prompts
ugly, deformed, disfigured
Avoid poor quality
blurry, low resolution, pixelated
Avoid technical issues
bad anatomy, bad hands, extra fingers
Fix common body issues
watermark, text, signature
Remove unwanted elements
worst quality, low quality
General quality filter
cropped, out of frame
Avoid composition issues
duplicate, cloned face
Prevent repetition
missing limbs, extra limbs
Anatomical corrections
nsfw, nude (if SFW desired)
Content filtering
monochrome, grayscale (unless wanted)
Color control
Img2Img Workflow
Denoising strength 0.3-0.5
Minor modifications
Denoising strength 0.5-0.7
Moderate changes
Denoising strength 0.7-0.9
Major transformation
Denoising strength 0.9+
Almost new image
Match resolution to input
Avoid artifacts
Sketch to detailed art
High denoising (0.7-0.8)
Photo enhancement
Low denoising (0.2-0.4)
Style transfer
Medium denoising (0.5-0.6)
Inpainting for edits
Selective region modification
Outpainting for expansion
Extend image borders
ControlNet
ControlNet Canny
Edge detection, precise composition
ControlNet Depth
Depth map guidance
ControlNet OpenPose
Pose skeleton control
ControlNet Scribble
Rough sketch to image
ControlNet Normal
Surface normal maps
ControlNet Seg
Semantic segmentation
ControlNet Lineart
Line art conversion
ControlNet MLSD
Straight line detection
Control weight 0.5-1.0
Adjust influence strength
Starting control step 0
When control begins
Ending control step 1.0
When control ends
Multiple ControlNets
Combine for complex control
Resolution & Aspect Ratios
512x512
SD 1.5 base resolution
768x768
Higher quality SD 1.5
1024x1024
SDXL base resolution
512x768 portrait
2:3 aspect ratio
768x512 landscape
3:2 aspect ratio
1024x576 widescreen
16:9 cinematic
Multiples of 64
Optimal for SD models
Hires fix for upscaling
Generate small, upscale big
Upscale denoising 0.4-0.6
Add detail when upscaling
Prompt Engineering Techniques
(word:1.2) emphasis
Increase word weight 20%
(word:0.8) de-emphasis
Decrease word weight
[word] alternation
Swap words during steps
{word|alternative} random
Random choice per generation
Comma separation
Distinct concepts
Order matters
Earlier words weighted more
Token limit 75
SD 1.5 prompt length limit
Token limit 150
SDXL allows longer prompts
Break into segments
Use AND for long prompts
Subject, style, quality
Prompt structure template
Model Selection
SD 1.5 base
General purpose, fast
SD 2.1
Better quality, slower
SDXL 1.0
Highest quality, requires more VRAM
Custom models (ckpt/safetensors)
Community fine-tuned models
LoRA (Low-Rank Adaptation)
Style/character additions
Textual Inversion
Embed specific concepts
VAE (Variational Autoencoder)
Affects color/quality
Realistic models
Photo-realistic results
Anime models
Illustration style
Artistic models
Painting/drawing styles
Batch & Variations
Batch count
Sequential generations
Batch size
Parallel generations (VRAM intensive)
Seed -1 for random
Different result each time
Fixed seed for consistency
Reproducible results
Variation seed
Similar but different results
Variation strength
How much to vary
Subseed for subtle changes
Minor variations
X/Y/Z plot
Compare parameters visually
Advanced Workflows
Txt2img → Img2img refinement
Two-stage generation
Low res → Hires fix → Img2img
Quality enhancement pipeline
ControlNet + img2img
Precise control + refinement
Inpainting for fixes
Correct specific areas
Multiple LoRAs
Combine styles/concepts
Prompt scheduling
Change prompt during generation
Regional prompting
Different prompts for areas
Composable Diffusion
AND for multiple concepts
Pro Tip:
Start with DPM++ 2M Karras sampler at 20-30 steps, CFG 7-8 for balanced results. Use negative prompts liberally! For faces and hands, try img2img refinement at 0.3-0.4 denoising. ControlNet OpenPose solves most pose problems. Always use hires fix for large images!