How do you lock a scene element to generate consistent multi-angle AI shots?

Lock one element of the scene as a fixed reference — a character sheet, a prop image, or a world plate — then ask the invideo agent for the angles you want around it. Once that anchor is in context, the agent generates wide, close, side, and reverse coverage off the same locked element without you re-prompting each shot.

Treat the locked element as the anchor and the camera as the orbit. The workflow is three steps, in order.

1. Build the anchor asset before any video generation. For a character, generate a 4-angle turnaround sheet (front, side, profile, back) plus a face close-up and a mid-angle close-up at 4K — Nano Banana or GPT-Image-2 handles the sheet, and you generate four options per asset and pick one before locking. For a prop, generate the prop in isolation (and remove props from characters' hands in their sheets, so the prop doesn't drift across angles). For a location, have the invideo agent scout real-world plates from the internet and pick the frames you want as the world reference. The principle running through all three: lock character sheets and environment references before a single video clip is generated — that is the step that prevents consistency problems across the rest of the film.

2. Anchor it in the agent's context, not in every prompt. invideo is an agentic video tool with the current image and video models — Recraft, Nano Banana, GPT-Image-2, Runway, Veo, Kling, Seedance 2.0 — sitting behind one agent that holds your project's context. Upload the locked sheet/prop/plate once and tell the agent to save it as the anchor for this scene: "I want you to deeply understand this and save it into context for further generations." Once locked, every shot the agent generates inherits that anchor — the invideo agent autonomously attaches the right reference image when it builds each new angle, instead of you re-uploading per shot.

3. Vary only the camera, hold everything else. Now ask for angles in plain directorial language: "hold on him in a wide, then a low-angle push-in, then the over-the-shoulder reverse." Lock one element and the agent extracts every angle — wide, close, side — without you requesting each individually. For reverse coverage, the agent applies art-director logic rather than mirroring: it surfaces what's undecided ("that near wall doesn't exist yet — what should it be?") and offers options before generating, so the reverse stays spatially consistent with the hero. Route the actual generation to Seedance 2.0 reference-to-video when you need the anchor carried into motion — it accepts character references and location references simultaneously, which is what makes multi-angle coverage of the same scene hold together. For the still-frame multi-angle pass, GPT-Image-2 or Nano Banana on the anchor sheet gives you the angle spread you select from.

A few things that make the lock actually hold across angles. Per-beat sheets when the anchor evolves: if a character picks up a trinket between angles, generate a separate sheet for that beat so the angle change doesn't erase the new element. Explicit inclusion/exclusion when you anchor — tell the agent what to take from the reference and what to ignore. And when a continuity error does appear in one angle, ask the invideo agent to inspect the character sheet rather than re-rolling the shot; it identifies the exact panel containing the error, fixes it at source, stores the corrected sheet, and every subsequent angle inherits the fix. Surgical edits, not slot-machine re-rolls.

As Hridaye, invideo's creative director, puts it: "Locking one element of the world causes the agent to automatically extract every angle — wide, close, side — without being asked." That is the whole technique in one line: one anchor, locked in context, then ask for the orbit.

Watch some of these to see what works for you:

Locking one element of the world causes the agent to automatically extract every angle — wide, close, side — without being asked.

— Hridaye, invideo's creative director