How do you use AI to expand a small studio set into a larger world?

You expand a small set by shooting your practical action framed tight enough to hide the set edges, then uploading screenshots of that footage into an AI agent as visual references and generating wide environment shots that place the same action in a larger world. One production turned a 6x6-foot studio pool into open ocean this way.

Start with the practical shot, not the prompt — AI extends physical production rather than replacing it. Build and shoot your set as usual, framing tight enough that the edges never appear: in one documented production, the emergence-from-water shot came from a studio pool only 6 ft by 6 ft, and the team separately built an 8x8-foot practical cave wall before adding AI effects on top. Physical constraints also shape your coverage plan — the actor in that pool could tread water for only one to two minutes per take, so the tight practical shots stayed short and the scale had to come from AI.

Next, anchor the AI to your footage. Upload screenshots or stills from your own shoot into the invideo agent as visual references before generating anything new, so color, contrast, and look match your practical material — invideo is an agentic video creation tool with the current generation models available, and it routes environment generations to models like Seedance 2.0, which produced the ocean and shoreline footage in this workflow. As the filmmaker behind it put it: "You definitely get so much more out of AI when you actually use your own footage as the references. You can make it look very close to something that looks real."

Then generate outward from the locked shot, letting story logic pick the environments. Ask for wide shots that place the same action in a bigger world — if your practical frame shows someone on hands and knees coming out of water, a shoreline wide is the shot that makes the cut feel continuous, not whichever generation looks most impressive. Keep the AI strictly on wide environment shots: pushing in close on a character doing any acting is where the footage reads as fake, so pull out rather than push in.

Iterate conversationally rather than one-shotting prompts — go back and forth, tell the invideo agent what you like and don't like, and pin a strong generation to request variations ("another one like it, but change X, Y, and Z"). The first shot teaches the look, tone, and effect; in the documented production, once that visual language was established, the second shot in the scene landed on the very first generation.

Finally, blend the two worlds in post. Match environmental continuity cues — the same haze in your generated exteriors as in your practical interiors — then color grade the AI footage to your practical material and layer sound design over it, which is the multiplier that makes generated inserts feel real. Since a single generation is rarely perfect end to end, splice the best moments from multiple generations into one polished sequence.

Watch some of these to see what works for you:

You definitely get so much more out of AI when you actually use your own footage as the references. You can make it look very close to something that looks real.

— Alex Arfaoui, filmmaker