Image-to-Image Generation

​

How It Works

1. Encodes your input image to latent space
2. Adds controlled noise (determined by denoise strength)
3. Denoises using your prompt as guidance
4. Decodes back to pixel space

• 0.0: No change (original image)
• 0.3-0.5: Subtle modifications, preserves composition
• 0.6-0.8: Significant changes, keeps general structure
• 0.9-1.0: Major transformation, minimal original influence

​

Basic Workflow

​

Complete Workflow JSON

{
  "load_image": {
    "class_type": "LoadImage",
    "inputs": {
      "image": "example.png"
    }
  },
  "checkpoint": {
    "class_type": "CheckpointLoaderSimple",
    "inputs": {
      "ckpt_name": "v1-5-pruned-emaonly.safetensors"
    }
  },
  "vae_encode": {
    "class_type": "VAEEncode",
    "inputs": {
      "pixels": ["load_image", 0],
      "vae": ["checkpoint", 2]
    }
  },
  "positive": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "clip": ["checkpoint", 1],
      "text": "watercolor painting, soft colors, artistic"
    }
  },
  "negative": {
    "class_type": "CLIPTextEncode",
    "inputs": {
      "clip": ["checkpoint", 1],
      "text": "blurry, distorted, low quality"
    }
  },
  "sampler": {
    "class_type": "KSampler",
    "inputs": {
      "model": ["checkpoint", 0],
      "positive": ["positive", 0],
      "negative": ["negative", 0],
      "latent_image": ["vae_encode", 0],
      "seed": 12345,
      "steps": 25,
      "cfg": 7.5,
      "sampler_name": "dpm++ 2m karras",
      "scheduler": "karras",
      "denoise": 0.65
    }
  },
  "vae_decode": {
    "class_type": "VAEDecode",
    "inputs": {
      "samples": ["sampler", 0],
      "vae": ["checkpoint", 2]
    }
  },
  "save": {
    "class_type": "SaveImage",
    "inputs": {
      "images": ["vae_decode", 0],
      "filename_prefix": "img2img"
    }
  }
}

​

Common Use Cases

​

Style Transfer

• Denoise: 0.5-0.7
• Prompt: Describe target style in detail
• CFG: 8-10 for strong style application

Positive: "anime style, cel shaded, vibrant colors, studio ghibli"
Negative: "realistic, photographic, 3d render"

​

Upscaling and Enhancement

1. Add ImageScale node before VAEEncode
2. Set target dimensions (e.g., 2x original)
3. Use low denoise (0.3-0.5)
4. Prompt for enhanced details:

Positive: "high resolution, sharp focus, detailed, 8k, masterpiece"
Denoise: 0.4

​

Sketch to Image

• Denoise: 0.7-0.9 (high transformation)
• Detailed descriptive prompt
• Higher CFG (9-12)

Positive: "professional character concept art, full color, detailed shading, 
fantasy armor, dramatic lighting"
Denoise: 0.85

​

Photo Restoration

• Denoise: 0.3-0.5 (preserve original)
• Steps: 30-40
• Sampler: ddim (deterministic)

Positive: "restored vintage photograph, natural colors, sharp details, 
high quality scan"
Negative: "damage, scratches, blur, grain, artifacts"
Denoise: 0.4

​

Advanced Techniques

​

Two-Pass Workflow (Hires Fix)

1. First pass: Generate at 512×512, denoise 1.0
2. Add LatentUpscale node to 2x size
3. Second pass: Same prompt, denoise 0.4-0.6
4. Results in higher quality than direct high-res generation

​

Variation Generation

1. Set denoise to 0.4-0.6
2. Use RepeatLatentBatch to duplicate encoded image
3. Each batch gets a different seed
4. Generate multiple variations in one run

​

Image Blending

1. Encode both images with VAEEncode
2. Use LatentBlend node (if available)
3. Set blend_factor (0.5 = equal mix)
4. Denoise the blended result

​

Denoise Strength Guide

​

Resolution Handling

​

Resizing Input

{
  "resize": {
    "class_type": "ImageScale",
    "inputs": {
      "image": ["load_image", 0],
      "width": 1024,
      "height": 1024,
      "upscale_method": "lanczos",
      "crop": "center"
    }
  }
}

• lanczos: Best for upscaling photos
• bicubic: Smooth results
• nearest-exact: Pixel art
• bilinear: Fast but softer

​

Aspect Ratio Preservation

{
  "width": 1024,
  "height": 0  // Automatically calculated
}

​

Sampler Recommendations

​

For Photo Editing

• ddim: Deterministic, preserves details
• euler: Fast, good for previews
• Steps: 20-30

​

For Artistic Transformation

• dpm++ 2m karras: Quality/speed balance
• euler_ancestral: More variation
• Steps: 25-35

​

For High Fidelity

• dpm++ sde karras: Best quality
• ddim: Consistent results
• Steps: 30-50

​

Troubleshooting

​

Too much change from original

• Lower denoise strength (try 0.4)
• Reduce CFG scale (try 6-7)
• Use ddim sampler for more consistency

​

Not enough change

• Increase denoise (try 0.7-0.8)
• Raise CFG scale (try 9-11)
• Be more specific in prompt
• Increase steps to 30-40

​

Artifacts or distortion

• Check if image resolution is compatible (multiples of 8)
• Try different VAE
• Lower denoise strength
• Reduce CFG scale

​

Out of memory

• Scale down input image
• Use VAEEncodeTiled for very large images
• Reduce batch size

​

Tips and Best Practices

1. Start conservative: Begin with denoise 0.5, adjust from there
2. Match model training: Use SD1.5 for 512px, SDXL for 1024px
3. Seed consistency: Lock seed when iterating on denoise/prompts
4. Prompt the original: Include aspects you want to preserve
5. Negative prompt matters: Be specific about what to avoid

​

Next Steps

• Learn Inpainting to edit specific regions
• Explore ControlNet for precise structural control
• Try combining img2img with LoRAs for enhanced results
• Experiment with different models optimized for specific tasks