How is AI used to extend practical film sets in a hybrid production workflow?

AI extends practical sets by anchoring generations to your own footage: shoot the action on a small physical build, upload stills from that shoot as visual references, then generate wide environment shots that place the same action in a larger world. In one documented production, a 6x6 ft studio pool became a full night ocean this way.

Build the practical piece first — AI extends physical production, it doesn't replace it. Construct only what the camera and actor actually touch: one documented hybrid production built an 8x8 ft cave wall and shot an emergence-from-water scene in a 6x6 ft studio pool too small to hide its own edges. Physical constraints define the split — the actor could tread water for only one to two minutes per take, so the practical camera covered tight action while AI supplied everything wider.

Upload stills from your own shoot as references. Before generating anything, screenshot your practical footage and feed those frames to the generator so color, contrast, and lighting are anchored to material that already exists. This is the single biggest realism lever in a hybrid workflow — reference-anchored generations match your plates far more closely than text-only prompts. invideo is an agentic video creation tool with the current video models available, and the invideo agent accepts these stills directly, then runs the generation as a back-and-forth conversation; in the documented production it produced the red-ocean and shoreline environments through Seedance 2.0 from exactly this kind of reference input.

Keep AI on wide environment shots and generate outward from your locked practical shot. Treat the practical footage as the fixed anchor and ask for wides that place the same action in a bigger world — a full shoreline around the pool shot, a cave exterior around the built wall. Pull out rather than push in: AI-generated characters break down in close-up acting shots, so anything with performance stays practical. Let story logic pick the AI shots, not visual impressiveness — the shoreline wide worked because the practical screenshot showed the actor on hands and knees coming out of water, so the two shots read as one continuous scene.

Establish the look on the first shot, then reuse it across the scene. Use your first generation to lock the look, tone, and effect; subsequent shots in the same scene come faster and more accurately because the session has internalized what you want — in the documented production, the second cave shot succeeded on the very first generation. Work iteratively: pin a generation you like and ask for variations ('create another one like it, but change X, Y, and Z'), and tell the invideo agent explicitly what you like and don't like — active feedback compounds across the session.

Blend everything in post. Expect partial value, not perfect clips: extract the best moments from multiple imperfect generations and splice them into one sequence. Color grade AI and practical footage together and add sound design — sound is the multiplier that makes generated inserts feel physical. Match environmental continuity cues, like carrying the same haze from an AI exterior into your practical interior, so transitions read as one location. For generated footage that looks artificial at normal speed, ramp it hard in your editor — one production pushed AI energy clips to 1,000%, 2,000%, even 5,000% in Premiere Pro and cut the fragments into a fast insert sequence that read as cinematic.

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