Text-to-Image Generation

​

Core Concepts

​

The Text-to-Image Pipeline

1. Text Encoding: Your prompt is converted to embeddings via CLIP
2. Latent Generation: Start with random noise in latent space
3. Denoising: The model iteratively removes noise guided by your prompt
4. Decoding: VAE converts the latent to a visible image

​

Key Nodes

• CheckpointLoaderSimple: Loads MODEL, CLIP, and VAE
• CLIPTextEncode: Converts text to conditioning
• EmptyLatentImage: Creates the initial noise tensor
• KSampler: Performs the denoising process
• VAEDecode: Converts latent to pixels
• SaveImage: Saves the final output

​

Basic Workflow

A serene mountain landscape at sunset, dramatic lighting, 
high detail, photorealistic, 8k resolution

blurry, low quality, distorted, ugly, bad composition

​

Advanced Prompting

​

Emphasis Syntax

(masterpiece:1.3)  // Increase emphasis by 30%
(background:0.7)   // Decrease emphasis by 30%
((very important)) // Nested = (very important:1.21)

​

Dynamic Prompts

A {red|blue|green} sports car in {rain|snow|sunshine}

​

Comments

A beautiful landscape // This is ignored
/* Multi-line
   comments also work */

​

Textual Inversion

embedding:my_style.pt, detailed portrait

​

Parameter Guide

​

Seed

• Purpose: Controls randomness
• Same seed + settings = identical output
• Tip: Lock seed when iterating on prompts

​

Steps

• Range: 15-50 typically
• 15-20: Fast preview quality
• 25-35: Production quality
• 40+: Diminishing returns

​

CFG Scale

• Range: 1-20
• 1-5: Creative, loose interpretation
• 7-9: Balanced (recommended)
• 10-15: Strict adherence to prompt
• 15+: May oversaturate or distort

​

Sampler Selection

• euler: Simple, fast, good for previews
• euler_ancestral: Adds variation between steps

• dpm++ 2m karras: Excellent quality/speed balance
• dpm++ sde karras: High quality, slower
• dpm++ 2m sde karras: Best quality, slowest

• ddim: Deterministic, good for img2img
• uni_pc: Fast convergence, experimental

​

Scheduler

• normal: Linear noise schedule
• karras: Better for DPM++ samplers
• exponential: Smoother transitions
• sgm_uniform: For SD3/SDXL

​

Workflow Example: High-Quality Portrait

{
  "checkpoint": {
    "class_type": "CheckpointLoaderSimple",
    "inputs": {
      "ckpt_name": "sd_xl_base_1.0.safetensors"
    }
  },
  "positive": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "clip": ["checkpoint", 1],
      "text": "professional portrait photograph, (soft studio lighting:1.2), detailed face, natural expression, bokeh background, Canon 5D, 85mm lens, f/1.8"
    }
  },
  "negative": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "clip": ["checkpoint", 1],
      "text": "blurry, distorted features, bad anatomy, oversaturated, artificial, plastic skin"
    }
  },
  "empty_latent": {
    "class_type": "EmptyLatentImage",
    "inputs": {
      "width": 1024,
      "height": 1024,
      "batch_size": 1
    }
  },
  "sampler": {
    "class_type": "KSampler",
    "inputs": {
      "model": ["checkpoint", 0],
      "positive": ["positive", 0],
      "negative": ["negative", 0],
      "latent_image": ["empty_latent", 0],
      "seed": 42,
      "steps": 30,
      "cfg": 8.0,
      "sampler_name": "dpm++ 2m karras",
      "scheduler": "karras",
      "denoise": 1.0
    }
  },
  "vae_decode": {
    "class_type": "VAEDecode",
    "inputs": {
      "samples": ["sampler", 0],
      "vae": ["checkpoint", 2]
    }
  },
  "save": {
    "class_type": "SaveImage",
    "inputs": {
      "images": ["vae_decode", 0],
      "filename_prefix": "portrait"
    }
  }
}

​

Batch Generation

1. Set batch_size in EmptyLatentImage to 4 (or higher)
2. Keep seed the same for subtle variations
3. Or use different seeds for diverse results

​

LoRA Integration

1. Add LoraLoader node
2. Connect MODEL and CLIP from checkpoint
3. Select LoRA file
4. Set strength_model (0.5-1.0 typical)
5. Set strength_clip (0.5-1.0 typical)
6. Connect LoraLoader outputs to KSampler

{
  "lora": {
    "class_type": "LoraLoader",
    "inputs": {
      "model": ["checkpoint", 0],
      "clip": ["checkpoint", 1],
      "lora_name": "add_detail.safetensors",
      "strength_model": 0.8,
      "strength_clip": 0.8
    }
  }
}

​

Troubleshooting

​

Black/blank images

• Check VAE compatibility with your model
• Try loading a separate VAE with VAELoader
• Reduce CFG scale

​

Low quality results

• Increase steps (try 30-40)
• Use better sampler (dpm++ 2m karras)
• Improve prompt detail and specificity
• Check if model supports your resolution

​

Out of memory

• Reduce image dimensions
• Lower batch size
• Use tiled VAE for very large images
• Enable CPU offloading: --lowvram flag

​

Next Steps

• Explore Image-to-Image for variations
• Learn Inpainting for editing
• Try ControlNet for precise control
• Experiment with different models and LoRAs