What is the mock-shot reference technique and when should you use it in AI video production?

When AI video models can't produce a specific camera angle from text prompts — especially POV shots and multi-character contact shots — physically act the shot out on your phone, then upload that footage as a reference video for the model to lock onto. The invideo agent routes that mock clip into Seedance 2.0 reference-to-video alongside your character sheets, so the spatial and motion logic carries through.

Use it the moment prompting stalls. POV shots and multi-character physical contact shots (ropes, props, bodies touching) break current video models faster than almost anything else, and no amount of rephrasing fixes them — text can't convey the exact eyeline, hand position, or camera path you need. A 5-second phone clip can.

The workflow runs in four steps:

1. Exhaust prompt iteration first, then switch modes. If two or three generations come back wrong on framing or POV geometry, stop rewriting the prompt. The signal isn't "prompt better" — it's "the model needs a spatial reference."

2. Shoot the mock on your phone. Stand in for the character, hold the phone at the eyeline you want, and act the move out — a walk-up, a hand reach, a head turn, whatever the shot demands. Keep it short (a few seconds), keep lighting roughly close to your scene, and frame it the way you want the final shot framed. You're feeding the model camera position, motion, and timing — not performance.

3. Upload it to the invideo agent with your locked references. Hand the mock clip to the invideo agent along with your character sheets and any environment plates. The agent attaches all of it to Seedance 2.0 reference-to-video, which ingests the mock's motion and spatial context plus the character/location references in one pass — something text-only or single-image-reference flows can't do.

4. Generate, then composite if needed. Expect about 3 generations per usable shot; roughly 40% of finished shots get stitched from 2+ generations, so treat the mock-reference output the same way — pull the strongest seconds from each attempt and combine them.

The same principle covers adjacent dead-ends. When an image model can't visualize a complex multi-character configuration from text, hand-sketch the arrangement on paper, photograph the drawing, and upload it as the reference instead — the invideo agent feeds it into the image model and generates the character sheet from there. Drawing and shooting are the two physical-world inputs that unblock model limitations.

A note on model choice: Seedance 2.0 reference-to-video is the strongest target for this technique because it accepts a video reference AND character/location references simultaneously, so the mock's motion fuses with your locked identity assets. Veo and Kling also accept video references and are worth routing to for different looks — the invideo agent holds all of these models, so you don't pick a platform per model; the agent picks the model per shot. Hridaye, invideo's creative director, frames the underlying posture this way: "a more efficient way to go about doing it is actually just using few world reference images and character sheets. Getting agent to upload that to see Dan's reference to video and then truly just prompting it like a director prompts his crew." The mock clip is one more reference you hand the crew.

Quick sanity check on what a usable mock looks like: clear framing, the camera motion you actually want in the final, decent light so the agent can read the geometry, and just the action — no dialogue, no clutter. You're giving the model a spatial diagram, not a performance.

Watch some of these to see what works for you:

The lesson for the day truly is that when the models get stuck you draw, you shoot, you bring your hands in and you get it done. And that's when agent one meets you there and takes it over the line.

— Hridaye, invideo's creative director