Both workflows follow the same broad pattern. First, break the video into a set of scenes, then plan the scene, write the code, render it, and iteratively refine the result.

For Remotion, the AI determines what resources the scene will need, whether that be images, generated assets, or prebuilt visual components. This part doesn't happen in Manim because its more for math based visualizations.

Re costs, both workflows have a series of presets that make the implementation less expensive and more consistent. The reduction in price for the Terra model series has been particularly helpful for us.