1. Workflow Overview

This workflow specializes in historical photo restoration with:

• AI Colorization: Convert B&W to natural colors

• HD Upscaling: 1.5x resolution boost (4K capable)

• Damage Repair: Auto-fix scratches/folds/facial flaws

Core Technologies:

• majicMIX realistic v7: Photorealistic portrait model

• ioclab_sd15_recolor ControlNet: Colorization-optimized

• BNK_CLIPTextEncodeAdvanced: Enhanced prompt encoding

2. Key Components

3. Critical Parameters

• Color Control:（Python）

  "ControlNetApplyAdvanced": [1, 0, 1]  # strength=1.0 (full process)

• Negative Prompt:（Python）

  "paintings,((monochrome)),((grayscale))..."  # force color output

• Sampler:（Python）

  "KSampler": ["dpmpp_2m", "karras", 30 steps, CFG=5]

4. Workflow Architecture

graph LR
  A[Input B&W Photo] --> B[1.5x Upscale]
  B --> C[ControlNet Colorize]
  C --> D[VAE Decode]
  D --> E[Comparison Output]

1. Preprocessing:

   • Upscale with nearest-exact algorithm

   • Preserve composition via VAEEncode

2. Colorization:

   • Dual prompts:

     • Positive: "8k wallpaper, best quality"

     • Negative: "monochrome, grayscale"

3. Output:

   • Side-by-side comparison with Image Comparer

5. I/O Specifications

• Input Requirements:

  • Format: JPG/PNG (600+ DPI recommended)

  • Resolution: Min 600x800px

  • Content: Front-facing portraits work best

• Outputs:

  • Resolution: Original x1.5 (e.g. 1000x1498→1500x2247)

  • Format: PNG (lossless)

  • Metadata included (view via PNGInfo)

6. Pro Tips

1. Hardware:

   • ≥8GB VRAM (for 1500x2247 processing)

   • Enable --xformers

2. Troubleshooting:

   • Over-saturation: Reduce ControlNet strength (1.0→0.8)

   • Face distortion: Add "bad anatomy" to negatives

   • Artifacts: Manual PS repair before reprocessing

3. Advanced:（Python）

   # More vibrant colors:
   Add "(vivid colors:1.3)" to positive
   # Partial B&W effect:
   Set ControlNet end=0.8 (default 1.0)