1. Workflow Overview

This workflow specializes in image relighting using LBM (Lighting Balance Model) to modify illumination conditions. Key features:

• Subject/background separation

• Physics-based lighting adjustment

• Automated masking & composition

2. Core Models

• LBM_relighting.safetensors: Main model (download from HuggingFace)

• TransparentBG: Background removal (base version)

• ImageResizeKJv2: Intelligent resizing algorithm

3. Critical Nodes

4. Processing Pipeline

Phase 1: Input Preparation

• Dual input channels:

  • Subject: image__00038_ (1).png (832x1216)

  • Background: 00141-338031962.png (1024x1024)

• Resolution normalization via ImageResizeKJv2

Phase 2: Background Processing

• TransparentBGSession+ creates removal session

• ImageRemoveBackground+ generates alpha mask

Phase 3: Relighting

1. Load LBM model (bf16 precision)

2. 20-step lighting adjustment with LBMSampler

3. Masked composition (ImageCompositeMasked)

Phase 4: Output

• Real-time preview (PreviewImage)

• Auto-saved results (path: LBM%date:yyyy-MM-dd%\Flux-)

5. I/O Specifications

• Input Requirements:

  • Subject: PNG with alpha channel recommended

  • Background: Resolution ≥1024x1024

• Output:

  • Relit composite image

  • Metadata preserved

6. Critical Notes

1. Manual Downloads Required:

   • LBM model: HuggingFace link

   • Save to /models/lbm/

2. VRAM Requirements:

   • Minimum 8GB (for 1024x1024)

   • bf16 precision recommended

3. Troubleshooting:

   • Check TransparentBG model if background removal fails

   • Adjust mask feathering for edge artifacts