How do you deconstruct a reference ad before recreating it with AI?

Deconstruct a reference ad in five passes before generating anything: detect every cut and extract one frame per scene, break the ad into timestamped beats with emotional logic, extract its camera language and pacing rules, transcribe the script and music structure, then lock a reviewed shot-by-shot breakdown table that governs all generation.

Upload the reference ad as an attachment and run the analysis before touching a single image or video credit. invideo is an agentic video creation tool, and the invideo agent reads uploaded videos directly — it watches the full ad and returns structured analysis you then review and lock.

1. Detect every cut and extract one frame per scene. Ask the invideo agent to identify the exact timestamp of every cut and pull one representative frame per scene. In one documented production, it detected 9 cuts automatically and extracted a frame for each — those frames become your visual template, so each scene can be recreated individually with structural fidelity to the original. The working prompt is direct: find where every cut happens, grab a frame from each scene, change only what must change.

2. Break the ad into named beats. Instruct the invideo agent to break the reference into beats with timestamps, locations, functions, and emotional logic — this maps why each moment works (hook tension, emotional payoff before product reveal, CTA register), not just what's on screen. Beats are what you preserve when recreating: one team held all 7 shot beats of a winning ad constant across recreated versions while swapping only cast, location, voiceover, and on-screen text, keeping the edit structure, music bed, and pacing untouched.

3. Extract camera language and pacing rules. Ask the invideo agent to pull the reference's production rules and organize them under structured headings — visual standard, cut rate, camera style, on-screen text treatment, energy, and ending pattern. Have it save these rules in the project context so every subsequent generation inherits them without re-prompting.

4. Transcribe the script and decode the music. The invideo agent transcribes the voiceover, identifies the structural elements worth keeping, and uses them as the script template for your new ad. This works even on foreign-language references: one production had the agent translate a Thai reference ad, decode its wordplay structure, and analyze its music so both voiceover and soundtrack could be adapted for a new market.

5. Lock a shot breakdown table before generating. Have the invideo agent compile everything into a shot breakdown table — shot number, duration, shot description, super text — then review and lock it before any image or video generation begins. A single breakdown pass in one production produced 10 shots across 5 sets. At the same stage, ask what stays the same versus what changes, and request an estimated generation breakdown (reference sheets, video clips, UI screens) so you know approximate spend before committing credits.

One caution on burned-in captions: if the reference ad has captions baked into the video, exclude the reference file from your generation prompts — models will reproduce the old captions in new outputs. Drive generation from the extracted breakdown, character sheet, and location sheet instead, and keep the reference strictly as analysis input.

Watch some of these to see what works for you:

Go through the reference ad, find where every cut happens, grab a frame from each scene. Redo all of them with the Nigerian woman, don't change anything else.

— invideo's creative team