How are people keeping the same AI-generated location consistent across multiple live-action camera angles?
I’m working on a cloud-only AI compositing workflow and trying to solve environmental continuity across a full edited scene.
The footage is shot normally with multiple setups: wides, mediums, close-ups, reverses, profiles, different camera positions, and different lenses.
The goal is to keep the original actors, performances, framing, and camera movement, but replace the practical location with a new AI-generated environment and make every shot feel like it was photographed in the same physical space.
My current setup:
- Nano Banana 2 / Pro in Google Flow for still compositing, environment replacement, and relighting
- Seedance 2.0 through Comfy Cloud for video
- MacBook Air, so the workflow needs to stay essentially cloud-based
The issue is that getting one shot to work is relatively easy. The difficult part is making an entire sequence maintain believable spatial continuity.
For example:
- 35mm wide looking down a hallway
- 85mm close-up from the opposite direction
- Side-profile two-shot
- Reverse angle
- Another wide from a different position
I need the generated environment to respond correctly to each camera position rather than just creating five vaguely similar hallways.
I’m trying to maintain:
- Architecture and room layout
- Doors, windows, furniture, columns, etc.
- Screen direction and spatial relationships
- Lighting direction
- Perspective
- Approximate lens behavior
- Depth and scale
- Subject relighting/contact with the environment
- Overall color and atmosphere
It doesn’t need to be geometrically perfect. It just needs enough faux spatial continuity that the audience believes all the shots were captured on the same set.
For anyone doing this with tools like Nano Banana, Seedance, Kling, Veo, etc., what workflow has been the most reliable?
Do you:
- Generate a master environment first?
- Create several canonical reference angles of the location?
- Build a rough floor plan or 3D blockout first?
- Generate clean environment plates before adding actors?
- Use previous generated shots as references for subsequent angles?
- Handle environment replacement and actor relighting as separate passes?
I’m especially interested in workflows that work without a powerful local GPU.
Basically: what is the best way to fake one coherent virtual location across multiple independently generated live-action shots?