In Which the Author Explores Machine Learning for Green Screen Keying and Relighting - Part 2

In Which the Author Explores Machine Learning for Green Screen Keying and Relighting - Part 2

image

This time, I’m pushing further into the world of machine learning, using 9 layers to rebuild and manipulate light.

Here’s the breakdown :

The 9 Layers of Relighting

1️⃣ Plate - Original footage.

2️⃣ Key - Isolate the subject.

3️⃣ Albedo - Pure color data.

4️⃣ Specular - Reflective highlights.

5️⃣ World Position - 3D space info.

6️⃣ Normal - Surface orientation.

7️⃣ Raw Relight - Unaltered light contribution.

8️⃣ Extracted Light - Isolated sources.

9️⃣ Relight - Final lighting adjustments.

The 5 Steps

1️⃣ Generate Passes

Use ComfyUI to generate Alpha, Albedo, Depth, Specular, and Normal passes.

2️⃣ Copycat for Temporal Stability

Train Nuke to create consistent image sequences.

3️⃣ World Position Creation

Use the depth pass to generate a world position pass in Nuke 3D.

4️⃣Convoluted Relighting

I animated a single green omni light in Nuke 3D. It sweeps from screen left to screen right and back again to demonstrate how lighting shifts dynamically.

5️⃣ Wait for Comments

Now, I sit back and prepare for the inevitable feedback on how I should be doing this.

This process isn’t just technical, it’s playful experimentation. Each layer offers a new opportunity to manipulate light, turning creativity into a tool to challenge expectations.

While traditional methods remain part of my teaching, experimenting with AI reveals exciting possibilities that extend beyond what old techniques can achieve.

Because in the end, VFX isn’t just about getting it right, it’s about having fun while breaking things and seeing what emerges. If you’re not a little messy, you’re not really doing it, are you? 🌟

P.S. Most of the values are exaggerated for demonstration purposes.