Micro-case · 5/5
Micro-case: a UGC testimonial in one scene
The format that still outperforms polish, and the one constraint that decides whether it survives three shots.
What you can do after this
Produce a three-shot UGC testimonial where the creator, room, and product stay continuous.
Nodes used
- reference
- generate_image
- generate_video
- extract_frame
- composer
Before you start
- Lesson two, for references.
- Lesson three, because you will re-run at least one keyframe before this is right.
In this lesson
- 0:00
Name the failure before you start
UGC ads break on three shots that read as three different people in three different rooms, rather than on bad rendering.
- 0:10
Write a scene bible first
One location, one outfit, one time of day, written down as fixed constants. Every prompt in the run repeats those words verbatim, which is what makes the shots belong to each other.
- 0:20
Wire identity and product into the same keyframe
A creator reference and a product reference both feed the generate image node. Order matters: identity first, product second, because referenceRoles is read in edge order.
- 0:40
Generate all three keyframes before any motion
Hook, demo, payoff, as stills side by side. This is the review gate, and it is the cheap one.
- 1:05
Fix drift at the still, never at the video
If a face has drifted, re-run that keyframe now. Motion is the expensive step; sending a broken keyframe into it pays twice for the same mistake.
- 1:25
Carry continuity with extract_frame
Pull the last frame of shot one and use it as the keyframe for shot two. The cut then lands inside a continuous room instead of jumping between two guesses at one.
- 1:35
Assemble with cuts
Three clips into the composer in order, transition set to cut. UGC does not dissolve, and a dissolve is the tell that a shoot was never there.
TranscriptThe same lines are burned into the video as subtitles.
- 0:00Micro-case: the UGC testimonial. The format that outperforms polish.
- 0:05The failure mode is obvious: three shots, three different people.
- 0:10So before generating anything, write a scene bible.
- 0:15One location, one outfit, one time of day. Non-negotiable.
- 0:20Add a creator reference — the face that has to repeat.
- 0:25Add the product reference next to it.
- 0:30Both wire into the same generate image node.
- 0:35referenceRoles is ordered: identity first, product second.
- 0:40Shot one: the hook. Creator to camera, product in hand.
- 0:45The prompt repeats the scene bible word for word. Every time.
- 0:50Shot two: the demo. Closer in, hands on the product.
- 0:55Shot three: the payoff. Back to the face, product forward.
- 1:00Three keyframes, one world. Check them side by side.
- 1:05If a face drifted, re-run that keyframe now.
- 1:10Motion is expensive. Keyframes are cheap. Fix it at the cheap step.
- 1:15Now each keyframe gets its own generate video node.
- 1:20Five seconds each, phone-camera framing in the prompt.
- 1:25For continuity, extract the last frame of shot one...
- 1:30...and use it as the keyframe for shot two.
- 1:35Wire all three clips into the composer, in order.
- 1:40Cut, not fade. UGC doesn't dissolve.
- 1:45Fifteen seconds, one creator, one room, one product.
- 1:50A testimonial ad nobody had to shoot.
Frequently asked questions
What is a scene bible?
A scene bible is the short list of constants every prompt in a run must repeat: the location, the outfit, the time of day, the camera treatment. It exists because models re-invent anything you leave unstated, and re-invention between shots is what makes an AI UGC ad look assembled.
Do I have to disclose that a testimonial is AI-generated?
Treat it as required. A synthetic creator is not a customer, so a generated testimonial should never be presented as a real person's experience, and most ad platforms now carry their own AI disclosure rules on top of that.
