Lore
Iterate by seed and prompt instead of simulation
The approach to fixing off results in the AI render workflow: regenerate, do not adjust a simulation. The model does not simulate physics, so splashes, secondary motion and small details are produced by sampling. When they look wrong, you sample again.
Procedure
- Details or physics look off, such as splashes: re-run with a different seed, or adjust the prompt (LLM-assisted prompt refinement and restyling).
- Only the ending is bad in a chained shot: resume from a later iteration (Iterative batching with overlapping start frames).
- Explore other looks: swap in alternate start frames, such as a mossy forest or an anime style. The layout and motion stay fixed because they come from the control passes.
- Start frame differs slightly from the control image: this is usually fine because the frame is only a reference. Changing the seed can also fix it.
Why it works
A generation takes minutes, so retrying is cheaper than setting up a physics simulation. The author's framing: 'You can explore 10 visual directions in the time it takes to set up one traditional render.' The cost is control: 'you're trading control for speed.' See AI rendering trade-off: speed for control (previs use case).