AI image + video production case study

Fix the frame.
Then make it move.

Most of the work happened after the first render. I compared versions and changed the direction when something broke. Sometimes that meant a wider image. Sometimes it meant splitting one video into two.

See the still revisions Watch the video iterations

My role: I planned the coverage and directed revisions. Nothing moved into the edit until I reviewed it.

Initial

Approved

These are the actual clips. They play muted.

The production decision

I fixed the still before I touched the video.

If a shirt looked like denim in the source image, the video carried that mistake forward. The same was true for faces, room layout, props, and framing. I approved the still first so the video rounds could focus on movement.

That saved some wasted attempts, but video brought its own problems. People moved too slowly. A pan went the wrong way. Two reference images fought each other. Those issues only showed up once the scene moved.

Still-image decisions

The still had to read in one second.

These weren't cosmetic tweaks. Each change made the action or setting easier to understand.

01 / Scene readability

The phone made the scene feel like a selfie.

I changed the action to laptop work and moved the camera back. The wider frame makes the laundry room and the work story clear right away.

Initial close-up of an attendee making a phone call in the prison laundry — Initial / unapprovedPhone-like close-up

Approved wider scene of an attendee working on a laptop in the prison laundry — Approved / canonicalWider laptop scene

02 / Composition

The close crop didn't read as a head count.

The first image focused on a few faces. Moving back put the officer and the full line in the same frame, so the action made sense without an explanation.

Initial close composition with a conference attendee, inmates, and an officer — Initial / unapprovedAction unclear

Approved wide view of a formal prison head count with the full line and officer visible — Approved / canonicalFull line and officer visible

03 / Story logic

The door bolt was within reach.

The image was supposed to communicate confinement, but the person inside could reach the release mechanism. I replaced it with a secured institutional gate.

Rejected prison-door image with the interior bolt and handle within reach — RejectedReachable release

Approved view of a locked institutional gate leading into the cell block — Approved / canonicalSecured threshold

04 / Wardrobe continuity

The scene worked. The sleeves didn't.

I kept the employee working from the bunk, along with the laptop and cell setting, then corrected the overshirt to match the approved short-sleeve wardrobe.

Superseded image of an employee working from a prison bunk in a long-sleeve overshirt — Revision neededLong sleeves

Approved image of an employee working from a prison bunk in a short-sleeve overshirt — Approved / canonicalShort sleeves

Fix it at the source

The shirts looked like denim, so I fixed the still first.

The first audience image had the wrong fabric. Asking the video model to repair it would have added another guess, so I corrected the image and used that version for the rerun.

Superseded keynote source still with denim-like blue shirts — Superseded sourceDenim-like texture

Approved keynote source still with powder-blue prison shirts — Approved sourcePowder-blue fabric

Video decisions

Then I had to solve the movement.

A clean still didn't guarantee a usable video. These are the actual attempts next to the versions I kept.

01 / Camera reaction

The camera needed a reason to move.

The straight walk felt mechanical. In the next version, he gestures toward the person filming. The camera jolts in response, which makes the moment feel connected instead of programmed.

SupersededStraight walk

Approved / canonicalGesture motivates camera reaction

02 / Gym background

More instructions made the movement less natural.

I had piled on a reveal, a second reference, and detailed directions for the exercises. The approved version stayed on the note-taker, gave him one glance up, and let the inmates in back keep working out.

Initial / unapprovedDetailed motion contract

Intermediate / unapprovedOverexpanded two-reference reveal

Approved / canonicalSimplified contained action

03 / Yard movement

Everyone in the yard needed to keep moving.

The first direction called out one dumbbell curl and froze too much around it. I replaced that with a broad instruction for everyone to keep exercising, plus one clear left pan.

Initial / unapprovedOver-constrained background

Approved / canonicalBroad exercise + left pan

04 / Whiteboard room

The second reference was hurting the transition.

I kept the whiteboard room as the only image and described the barred exit, hallway, and guards in words. That produced a cleaner move than supplying both endpoints.

SupersededTwo endpoint images

Approved / canonicalOne reference + described reveal

05 / Shuttle reveal

One long reveal kept breaking, so I split it.

The early version panned the wrong way and the second reference pulled the scene into a different shuttle interior. After many attempts, I stopped asking one clip to handle the whole move. Two short clips worked. This family had 19 completed generations and one failed render.

Initial / unapprovedWrong direction + reference interference

Approved / canonicalPart one

Approved / canonicalPart two

06 / Reading pace

One version dragged. The next rushed.

The seven-second version used fewer actions. One page turn, then a natural glance up. That gave the moment enough time without making it feel slow.

SupersededToo slow

RejectedToo fast

Approved / canonicalCalibrated

07 / First frame

The file rendered, but it started on the wrong frame.

I gave the rerun one more second and a simpler action. Because both changed, I can't claim which one fixed it.

RejectedOpening mismatch

ApprovedSimplified rerun

08 / Cafeteria reveal

The wider reveal worked, but the transition didn't.

The second reference gave us the tray placement, room reveal, and whisper. The editor still had to cut around the weak part of the transition.

InitialOne reference

Approved coverageEdit required

09 / Shirt movement

The still had short sleeves. The video didn't.

The first animation treated the overshirt like a long-sleeved garment. I kept the same source image and corrected the dressing action.

SupersededLong-sleeve behavior

ApprovedShort-sleeve correction

10 / Shirt fabric

The video copied the denim problem from the still.

I corrected the image first, then reran the scene. The rerun also changed how the source was supplied, so I don't treat this as a clean one-variable test.

Previously approved, superseded

Approved / canonical

How to read the labels

A finished file still needed a decision.

I labeled every version after reviewing it.

Initial / unapproved: Comparison only
Rejected: Reviewed and declined
Superseded: Replaced by a stronger version
Alternate: Valid, but not canonical
Approved / canonical: Selected to move forward
Experimental: Used to learn
Unresolved: Not claimed as approved

Evidence reviewed

The archive shows how much review this took.

These are snapshots from the files I recovered. They aren't a complete attempt count or a cost calculation.

30+selected stills in the archived snapshot

150+retired or rejected still variants

41distinct recovered videos reviewed

What I did

I watched each version, wrote down what wasn't working, and changed the direction. Only the images and clips that held up went into the editor's library.