The AI Filmmaking Pipeline
Editing locations
Editing is not a rescue step — it's part of developing a good input. Almost every good location still needs work: swapping details, removing clutter, color-correcting to your project. The skill is doing it without wrecking what already works.
Three models, run side by side
You don't pick one edit model — you run them in parallel and take the best result for each change.
| Model | Reach for it when |
|---|---|
| GPT Image 2 | Text, mathematically precise detail on complex objects |
| Nano Banana Pro | Most edits that don't need GPT-level precision |
| Seedream 4.5 | Run in parallel batches — rarely slops, but may miss the ask, so fire more tries |
Here's a real pass on a cafeteria location — wipe between the raw generation and the finished frame, then read what each tool contributed:
Nano Banana Pro cleaned the grime off the red metal, fixed the fused stools, and added trays and floor litter. GPT Image 2 swapped in the lamps, the TVs, and the menu boards, with readable text. Photoshop softened textures, removed the objects in the back corridor, unified the color, and composited every edit onto the original.
Every edit degrades the frame
The trap: an edit model never touches only what you asked for — it quietly re-renders the whole image. Watch one forest plate degrade, one edit at a time:
Every detail you add or remove gets masked onto the ORIGINAL in Photoshop — change only the patch that changed, leave the rest untouched. Never re-edit an edited image.
Lock in the rules
Edit the original
Composite each changed patch onto the original. Never stack edits — every full-frame rerender degrades untouched regions.
Run models in parallel
GPT Image 2 for text and precise detail, Nano Banana Pro for most edits, Seedream 4.5 in parallel batches.
Never re-edit an edit
A second pass re-renders the whole frame again, compounding slop. Always mask the change onto the master file.