Image-to-Image Edit

Transform existing images using AI diffusion models with text guidance. This skill enables powerful image editing capabilities including style transfer, content modification, inpainting, and outpainting.

Overview

Image-to-image (img2img) generation takes an existing image as input and transforms it based on a text prompt while preserving some characteristics of the original. The strength parameter controls how much the image changes - low values (0.3-0.5) preserve more of the original, while high values (0.7-1.0) allow more dramatic transformations.

Parameters

Required Parameters

Parameter	Type	Description
image_url	string	URL or local path to the source image. Supports PNG, JPG, WEBP formats. Must be publicly accessible URL or local file path.
prompt	string	Text description of the desired transformation. Be specific about style, content, and mood. Example: "transform into oil painting style, dramatic lighting, renaissance art"

Optional Parameters

Parameter	Type	Default	Description
negative_prompt	string	"blurry, low quality, distorted, watermark"	What to avoid in the generation. Use to exclude unwanted elements.
strength	float	0.75	Transformation intensity (0.0-1.0). Lower = preserve original, higher = more change. Recommended: 0.3-0.5 for subtle edits, 0.6-0.8 for style transfers, 0.9+ for near-complete rewrites.
seed	integer	random	Reproducibility seed. Same seed + same parameters = same output. Use -1 for random.
steps	integer	25	Inference steps (15-50). More steps = higher quality but slower. 20-30 is usually sufficient.
cfg_scale	float	7.5	Classifier-free guidance scale (1-20). Higher = more prompt adherence, lower = more creative freedom.
width	integer	source width	Output width in pixels. Leave empty to match source.
height	integer	source height	Output height in pixels. Leave empty to match source.
sampler	string	"dpmpp_3m_sde"	Sampling algorithm. Options: euler_a, dpmpp_3m_sde, dpmpp_2m, ddim, uni_pc, lms. DPM++ samplers recommended for quality.
scheduler	string	"karras"	Noise scheduler. Options: karras, normal, sgm_uniform, simple. Karras generally produces best results.
model	string	system default	Model checkpoint name. Examples: sdXL_v10, realisticVision_v20, dreamshaper_8.
mask_url	string	null	URL to mask image for inpainting. White areas = inpaint, black areas = preserve. PNG with alpha channel supported.
inpaint	boolean	false	Enable inpainting mode. Requires mask_url. Only modifies masked regions.
outpaint	boolean	false	Enable outpainting mode. Extends image beyond original boundaries.
outpaint_direction	string	"right"	Direction to extend: left, right, up, down, or combinations like "left,right,up".
batch_size	integer	1	Number of variations to generate. Each gets a different seed.
backend	string	"auto"	Which API to use: comfyui, automatic1111, stability, local. Auto-detects available backend.

Returns

{
  "image_url": "https://storage.example.com/generated/image_12345.png",
  "seed": 123456789,
  "metadata": {
    "prompt": "oil painting style, dramatic lighting",
    "negative_prompt": "blurry, low quality",
    "strength": 0.75,
    "steps": 25,
    "cfg_scale": 7.5,
    "sampler": "dpmpp_3m_sde",
    "model": "sdXL_v10"
  },
  "processing_time": 4.2
}

Usage Examples

Basic Style Transfer

// Transform photo into oil painting
await imageToImageEdit({
  image_url: "https://example.com/photo.jpg",
  prompt: "oil painting style, impressionist, visible brush strokes, vibrant colors",
  strength: 0.7
});

Subtle Enhancement

// Enhance photo quality while preserving content
await imageToImageEdit({
  image_url: "https://example.com/portrait.jpg",
  prompt: "professional photography, studio lighting, high detail, sharp focus",
  strength: 0.35,
  steps: 30,
  cfg_scale: 8
});

Inpainting (Edit Specific Areas)

// Replace object in image
await imageToImageEdit({
  image_url: "https://example.com/room.jpg",
  mask_url: "https://example.com/room_mask.png",
  prompt: "modern furniture, minimalist sofa, coffee table",
  inpaint: true,
  strength: 0.8
});

Outpainting (Extend Image)

// Extend image to wider aspect ratio
await imageToImageEdit({
  image_url: "https://example.com/landscape.jpg",
  prompt: "continue the landscape, mountains, sky, clouds",
  outpaint: true,
  outpaint_direction: "left,right",
  width: 1920,
  height: 1080
});

Batch Variations

// Generate multiple style variations
await imageToImageEdit({
  image_url: "https://example.com/portrait.jpg",
  prompt: "cyberpunk style, neon lights, futuristic",
  strength: 0.65,
  batch_size: 4
});
// Returns 4 images with seeds: seed, seed+1, seed+2, seed+3

Reproducible Results

// Same seed = same result
await imageToImageEdit({
  image_url: "https://example.com/photo.jpg",
  prompt: "watercolor painting style",
  strength: 0.6,
  seed: 42  // Always produces same output
});

Backend Implementations

ComfyUI Backend

ComfyUI uses a node-based workflow system. The img2img workflow:

1. LoadImage - Loads source image from URL or file
2. VAEEncode - Encodes image to latent space
3. CLIPTextEncode - Encodes positive/negative prompts
4. KSampler - Denoises latents with guidance
5. VAEDecode - Decodes latents back to image
6. SaveImage - Outputs final result

Key parameters in ComfyUI:

• denoise = strength (0.0-1.0)
• sampler_name = sampler algorithm
• scheduler = noise schedule

Automatic1111 Backend

Automatic1111's img2img API accepts:

{
  "init_images": ["base64_encoded_image"],
  "prompt": "transformation description",
  "negative_prompt": "what to avoid",
  "denoising_strength": 0.75,
  "steps": 25,
  "cfg_scale": 7.5,
  "seed": -1,
  "sampler_name": "DPM++ 3M SDE",
  "width": 1024,
  "height": 1024
}

Stability AI API

Stability AI's image-to-image endpoint:

{
  "init_image": "base64_or_url",
  "text_prompts": [
    {"text": "transformation description", "weight": 1},
    {"text": "negative prompt", "weight": -1}
  ],
  "image_strength": 0.75,
  "steps": 25,
  "cfg_scale": 7.5,
  "seed": 0
}

Strength Guidelines

Strength	Use Case	Example
0.2-0.4	Subtle fixes, upscaling, detail enhancement	"sharpen details, fix lighting"
0.4-0.6	Moderate changes, style enhancement	"add dramatic lighting, enhance colors"
0.6-0.8	Style transfer, significant changes	"transform into oil painting, cyberpunk style"
0.8-0.95	Major transformation, near-complete rewrite	"completely different scene, same composition"
0.95-1.0	Almost text-to-image, minimal original	"use as rough layout guide"

Prompt Engineering Tips

Be Specific About Style

❌ "make it artistic" ✅ "oil painting style, impressionist, visible brush strokes, warm color palette, dramatic lighting"

Describe Content Changes Clearly

❌ "change the background" ✅ "replace background with sunset beach scene, golden hour lighting, ocean waves"

Use Negative Prompts Effectively

negative_prompt: "blurry, low quality, distorted faces, extra limbs, watermark, text, signature"

Preserve Important Elements

prompt: "style as cyberpunk art, neon lights, futuristic city background, preserve the person's face and pose"

Common Use Cases

1. Photo to Art Style

{
  image_url: "photo.jpg",
  prompt: "oil painting style, renaissance art, dramatic chiaroscuro lighting, museum quality",
  strength: 0.7
}

2. Background Replacement

{
  image_url: "portrait.jpg",
  prompt: "professional studio background, gradient lighting, corporate headshot style",
  strength: 0.5,
  negative_prompt: "distracting elements, cluttered background"
}

3. Object Removal (Inpainting)

{
  image_url: "photo.jpg",
  mask_url: "mask.png",  // White over object to remove
  prompt: "clean background, seamless fill, natural continuation",
  inpaint: true,
  strength: 0.85
}

4. Color Grading

{
  image_url: "photo.jpg",
  prompt: "cinematic color grading, teal and orange, film look, Kodak Portra 400",
  strength: 0.3
}

5. Upscaling with Enhancement

{
  image_url: "low_res.jpg",
  prompt: "high resolution, sharp details, 4K quality, professional photography",
  strength: 0.4,
  width: 2048,
  height: 2048
}

6. Sketch to Image

{
  image_url: "sketch.png",
  prompt: "photorealistic rendering, detailed textures, professional photography",
  strength: 0.85
}

7. Day to Night

{
  image_url: "daytime_photo.jpg",
  prompt: "night scene, moonlight, street lamps, stars in sky, dark blue atmosphere",
  strength: 0.65
}

8. Season Change

{
  image_url: "summer_photo.jpg",
  prompt: "winter scene, snow covered ground, bare trees, overcast sky, cold atmosphere",
  strength: 0.7
}

Error Handling

Common Errors

Error	Cause	Solution
image_too_large	Image exceeds max dimensions	Resize to under 2048x2048
invalid_url	Image URL not accessible	Use public URL or local path
timeout	Generation took too long	Reduce steps or use faster sampler
out_of_memory	GPU memory exceeded	Reduce resolution or batch size
invalid_mask	Mask format incorrect	Use PNG with white/black regions

Retry Strategy

// Implement exponential backoff
async function robustImg2Img(params, maxRetries = 3) {
  for (let i = 0; i < maxRetries; i++) {
    try {
      return await imageToImageEdit(params);
    } catch (error) {
      if (i === maxRetries - 1) throw error;
      await new Promise(r => setTimeout(r, 1000 * Math.pow(2, i)));
    }
  }
}

Performance Optimization

Faster Generation

• Use euler_a sampler for speed (15-20 steps sufficient)
• Reduce steps to 15-20 for quick previews
• Lower resolution for iteration, upscale final
• Use smaller models (SD 1.5 vs SDXL)

Higher Quality

• Use dpmpp_3m_sde sampler with karras scheduler
• Increase steps to 30-50 for final output
• Use SDXL or fine-tuned models
• Higher cfg_scale (8-12) for prompt adherence

Memory Efficiency

• Process one image at a time
• Use tiled VAE for large images
• Enable memory-efficient attention
• Clear cache between generations

Integration with TaskingBot

This skill integrates with TaskingBot's image generation pipeline:

// Use with generate_image for comparison
const original = await generate_image({prompt: "landscape"});
const transformed = await imageToImageEdit({
  image_url: original.image_url,
  prompt: "oil painting style",
  strength: 0.7
});

Step-by-Step Instructions

Basic Image Editing Workflow

1. Prepare Source Image

   • Ensure image is accessible via public URL or local path
   • Recommended size: 512-1024px (will be resized automatically)
   • Supported formats: PNG, JPG, WEBP

2. Craft Your Prompt

   • Describe the desired transformation clearly
   • Include style, mood, lighting, and content changes
   • Use negative prompts to exclude unwanted elements

3. Set Strength Appropriately

   • 0.3-0.5: Subtle changes, enhancement
   • 0.5-0.7: Moderate style transfer
   • 0.7-0.9: Major transformation

4. Generate and Iterate

   • Start with default parameters
   • Adjust strength if too much/little change
   • Use same seed for reproducibility
   • Generate multiple variations with batch_size

5. Post-Process

   • Download result from returned URL
   • Use enhance_image for upscaling if needed
   • Use edit_image for additional refinements

Inpainting Workflow

1. Create Mask Image

   • Use image editing software (Photoshop, GIMP)
   • White (#FFFFFF) = areas to change
   • Black (#000000) = areas to preserve
   • Save as PNG with same dimensions as source

2. Configure Inpainting

   {
     image_url: "source.png",
     mask_url: "mask.png",
     prompt: "what should appear in masked areas",
     inpaint: true,
     strength: 0.8
   }
   

3. Best Practices

   • Mask slightly beyond the area you want to change
   • Use higher strength (0.8-0.95) for complete replacement
   • Describe the full scene context in prompt

Outpainting Workflow

1. Set Target Dimensions

   • Original: 1024x1024
   • Target: 1920x1080 (extends left/right)

2. Configure Outpainting

   {
     image_url: "1024x1024_image.jpg",
     prompt: "continue the scene naturally",
     outpaint: true,
     outpaint_direction: "left,right",
     width: 1920,
     height: 1080
   }
   

3. Direction Options

   • "left" - extend left side
   • "right" - extend right side
   • "up" - extend top
   • "down" - extend bottom
   • "left,right,up" - multiple directions

API Reference

Endpoint

POST /api/image-to-image

Request Headers

Content-Type: application/json
Authorization: Bearer <api_key>

Request Body

{
  "image_url": "https://example.com/image.jpg",
  "prompt": "transformation description",
  "negative_prompt": "what to avoid",
  "strength": 0.75,
  "seed": -1,
  "steps": 25,
  "cfg_scale": 7.5,
  "width": 1024,
  "height": 1024,
  "sampler": "dpmpp_3m_sde",
  "scheduler": "karras",
  "model": "sdXL_v10",
  "mask_url": null,
  "inpaint": false,
  "outpaint": false,
  "batch_size": 1
}

Response

{
  "success": true,
  "image_url": "https://storage.example.com/output.png",
  "seed": 123456789,
  "metadata": {
    "prompt": "transformation description",
    "strength": 0.75,
    "steps": 25,
    "processing_time_ms": 4200
  }
}

Best Practices Summary

1. Start Conservative: Begin with strength 0.5-0.6, adjust based on results
2. Use Negative Prompts: Always exclude "blurry, low quality, distorted"
3. Match Dimensions: Keep aspect ratio similar for best results
4. Iterate with Seeds: Note successful seeds for reproducibility
5. Batch for Variations: Use batch_size to explore options efficiently
6. Quality vs Speed: More steps = better quality, but slower
7. Prompt Specificity: Detailed prompts yield better results
8. Test Different Samplers: DPM++ samplers generally best for img2img

Troubleshooting

Image Not Changing Enough

• Increase strength (try 0.8-0.9)
• Make prompt more specific
• Increase cfg_scale (try 10-12)

Image Changing Too Much

• Decrease strength (try 0.3-0.5)
• Add preservation hints to prompt
• Use lower cfg_scale (try 5-7)

Artifacts or Distortions

• Increase steps (try 30-40)
• Try different sampler (dpmpp_3m_sde)
• Add "artifacts, distortion" to negative prompt

Slow Generation

• Reduce steps to 15-20
• Use euler_a sampler
• Lower resolution
• Reduce batch_size

See Also

• generate_image - Generate images from text prompts
• edit_image - Simple image editing (crop, filter, etc.)
• enhance_image - Upscale and improve image quality
• generate_3d_from_image - Convert image to 3D model

---

Version: 1.0.0 Author: TaskingBot Created: 2026-03-20 Tags: image, generation, editing, img2img, inpainting, outpainting, style-transfer, diffusion