AI Model Development for Nuke Compositing Software: From Scratch vs Fine-tuning

I want to develop an AI model for the Nuke Compositing Program. Initially, I’m aiming for a question-answer based system, and I’m currently collecting both synthetic and natural datasets.

I’m considering two approaches:

  1. Developing a model from scratch
  2. Taking an existing model and fine-tuning it

In later stages, I plan to have the model write code and perform node manipulations.

Which approach would you recommend? I would appreciate if you could share your experiences and suggestions.