Model Types
| Type | Size | Use Case | |------|------|----------| | SD1.5 | ~2GB | Fast, wide LoRA support, 512x512 | | SDXL | ~7GB | Higher quality, better prompts, 1024x1024 | | SD3 | ~7GB | Latest, improved text rendering | | Flux | ~12GB | Photorealism, very detailed | | Pony | ~7GB | SDXL fine-tune, stylized/anime |
Samplers
| Sampler | Speed | Quality | Use | |---------|-------|---------|-----| | Euler | Fast | Good | Quick tests | | DPM++ 2M Karras | Medium | Excellent | Default for quality | | DPM++ SDE Karras | Slow | Best | High detail | | DDIM | Medium | Good | Reproducible results | | LCM | Very Fast | Decent | Real-time generation |
Key Settings
CFG Scale: 7 (default), 3-5 (creative), 10+ (strict prompt adherence)
Steps: 20 (fast), 30 (standard), 40-50 (quality)
Denoising: 1.0 (txt2img), 0.3-0.5 (light img2img), 0.6-0.8 (heavy edit)
Clip Skip: 1-2 (SD1.5), 1 (SDXL - rarely change)Negative Prompt (Universal)
worst quality, low quality, lowres, blurry, ugly, deformed,
bad anatomy, bad hands, extra fingers, missing fingers,
watermark, signature, text, logo, jpeg artifactsHires Fix
Always generate at native resolution, then upscale:
- Base: 512x768 (SD1.5) or 1024x1152 (SDXL)
- Hires: 1.5-2x scale, 0.4-0.6 denoise
- Upscaler: 4x-UltraSharp or ESRGAN
Common Workflows
Text-to-image: Load Checkpoint → Empty Latent → CLIP Text Encode → KSampler → VAE Decode → Save
Image-to-image: Load Image → VAE Encode → [same as txt2img with denoise < 1.0]
Inpainting: Load image + mask → VAE Encode (masked) → KSampler (denoise 0.8-1.0) → VAE Decode
LoRA Usage
(lora:character-name:0.8) ← Include in prompt
(lora:style-name:0.6-0.8) ← Style LoRAs
(lora:concept-name:0.4-0.7) ← Concept/object LoRAsWeights above 1.0 tend to overpower the base model. Start at 0.6 and adjust.
VRAM Tips
--medvram : Required for 8GB cards with SDXL
--lowvram : For 4-6GB cards (slower but works)
--novram : For <4GB (falls back to CPU)
--always-offload: Prevents OOM on multi-model workflows