nuke-tiling-tool

Nuke Auto-Tiling Tool for ML Processing

Version: 1.0.0 Last Updated: 2026-01-21 Status: Production-ready Dependencies: Nuke 13.2v8+, auto_tile_processor.py

---

Purpose

Automatically tiles large images for processing through ML nodes (ViTMatte, etc.) that work best with smaller image sizes (1K-2K), then seamlessly blends the results back together using expression-based gradient masks.

When to Use This Skill

Trigger phrases:

• "tile image for ML processing"
• "split large image into tiles"
• "process image through ViTMatte in tiles"
• "auto-tile for machine learning"
• "create tiling setup"

Use cases:

• Processing large plates (4K, 5K, 8K) through ML nodes optimized for 1K-2K
• ViTMatte alpha generation on high-resolution images
• Any GPU-intensive node that performs better on smaller images
• Avoiding memory issues with ML inference

What It Does

Automated Workflow

Input Image (e.g., 5760x5760)
    v
Dot distribution (horizontal row)
    v
Auto-calculate grid (e.g., 3x3 for 2K tiles)
    v
For each tile (vertical stack):
    Transform (translate to position)
    -> Reformat (crop to tile size)
    -> NoOp Placeholder (orange - swap for ML node)
    -> Expression (gradient mask -> mask channel)
    -> Premult (multiply by mask.a)
    -> InverseTransform (restore position)
    v
Merge all tiles (plus operation, rgba output)
    v
Final Reformat (back to original dimensions)
    v
Output (full resolution, seamless blending)

Key Features

1. Auto-calculates grid dimensions

   • 5760x5760 with 2K tiles -> 3x3 grid
   • 8192x8192 with 2K tiles -> 5x5 grid
   • Non-square images supported

2. Expression-based gradient masks

   • Smooth falloff at tile edges (no blur, preserves detail)
   • 128px overlap with automatic blending
   • Edge/corner tiles handled correctly

3. Flexible tile sizes

   • 1K (1024x1024)
   • 2K (2048x2048)

4. Clean node organization

   • Horizontal tile layout (Nuke-standard)
   • Vertical node stacks per tile
   • Clear naming (Transform_Tile_X_Y, etc.)
   • Orange NoOp placeholders for ML node insertion
   • Dynamic positioning from input node location