How do you generate a shot/reverse-shot pair with consistent AI video?

Generate the hero shot first, then request the compositionally opposite angle in the same session — the scene's spatial geography carries over, so the reverse can be reconstructed without a new reference image. Consistency on both halves comes from locked character sheets and a locked location reference routed through Seedance 2.0 reference-to-video, not from hand-matching two prompts. This works well inside the invideo agent, which retains that session context across both requests.

invideo is an agentic video creation tool with all the current video models available, so the same invideo agent session can carry one scene's geography across both angles of the pair. The workflow runs in four steps.

1. Lock shared references before generating either angle. Both shots must inherit the same faces, wardrobe, and set, so build character sheets with multiple angles plus face close-ups — close-up panels keep small details like scars and accessories consistent across models — and lock a location reference. One documented production generated 4-angle character sheets at 4K and needed about 5 generations to lock one character at ~$9.78; another locked character sheets and environment references by generating 4 options per asset and picking the best before any video generation.

2. Generate the hero shot, then request the reverse in the same session. Once the hero angle is approved, immediately ask the invideo agent for the compositionally opposite angle in the same conversation — it uses the geography established by prior shots rather than starting from zero. In one documented production, the invideo agent reconstructed a precise reverse angle with no reference image supplied, using only the spatial logic of earlier shots; in another, a complex top-down hero shot landed on the first attempt and the matched opposite angle followed in the same session. Specify which side of the action line the camera sits on in both requests so eyelines stay opposed and the pair cuts together.

3. Resolve the reverse environment before generating. The reverse direction usually reveals set space you never designed, so instruct the invideo agent to apply art director logic: it surfaces the undecided elements — in one session, it asked, "Reverse on Marcus — what's behind him? That near wall doesn't exist yet. What should it be?" — and presents narrative-loaded options. Choosing one before generation prevents the model from inventing a wall that contradicts the hero shot.

4. Route both angles through Seedance 2.0 reference-to-video with the same reference bundle. Reference-to-video accepts character references and location references simultaneously — extend cannot — so both halves of the pair pull from identical visual anchors. You can also attach the approved hero clip itself as a video reference so camera language, lighting, and atmosphere carry over into the reverse.

Expect some iteration on the second angle: one production averaged 3 generations per usable shot, and over-the-shoulder framings are a documented weak point of Nano Banana — if a framing stalls, the invideo agent can reroute the shot to a different model without you re-engineering the pipeline.

Watch some of these to see what works for you:

it delivered a precise reverse angle without a single reference from me. That's pretty insane, right?

— invideo's creative team, documenting an AI short film production