AI Video line · stop 12 of 14 · 14 min · members
Deciding the frame before you generate, and why reframing afterwards costs more than it looks.
Free with an account
Membership is free: an account opens all 86 script pages. The Lab, Studio Canvas and the paid guides need the $99 pass, paid once. Already signed in on this browser? The page opens by itself.
The mistake
The model composed for the frame you asked for, and you are discarding that.
It seems efficient to generate one wide frame and crop it to whatever each platform needs. In practice the wide frame was composed as a wide frame — the subject placed, the balance struck, the empty space distributed for that shape.
Crop it to a tall format and the composition is gone. The subject sits wrongly, the negative space is in the wrong place, and the result reads as a crop rather than as a frame.
Generating for the ratio you need is not more expensive. It is the same generation, asked for correctly.
What changes with ratio
Models compose differently for different shapes, because their training did.
Ask for a wide frame and you tend to get environment, context and lateral balance. Ask for a tall one and you get proximity, a closer subject and vertical layering.
This is not a setting applied afterwards; it changes what the model produces. A portrait generation of the same prompt is frequently a different shot, not the same shot in a different container.
Which means the ratio is a creative decision that belongs at the start, alongside the subject and the treatment.
Deciding
Two questions that settle it immediately.
Where — a feed, a website header, a print piece, a presentation. Each has a shape it wants and fighting it costs attention.
In what state — full screen, in a grid of thumbnails, small on a card. This affects composition more than ratio does: a frame destined for a thumbnail needs one clear idea, whatever its shape.
Where a piece genuinely needs several ratios, generate for each rather than deriving. Two generations of the same prompt at two ratios cost less than one generation plus an hour of compositing.
Leaving room
Type, logos and platform interface all occupy the frame.
Almost every delivered frame has something over it. A headline, a brand mark, captions, or the platform's own interface eating the top and bottom of a vertical video.
Ask for that space explicitly — 'the lower third is empty', 'the subject sits in the left half' — and you get a frame that can be laid out. Ask for a beautifully balanced composition and you get one that has nowhere for the words to go.
This is the most common late-stage failure in generative work: forty good frames, none with room for the copy, discovered at layout.
Multiple deliverables
Where you must derive, plan the derivation from the start.
When several ratios are genuinely required and separate generations are not practical, generate wider than the widest deliverable and compose knowing it will be cut.
That means keeping the subject and the essential content inside the tallest crop, and treating the outer areas as expendable. It is the same discipline as shooting for multiple aspect ratios with a camera, and it works for the same reason.
Check every crop before delivery rather than trusting the master. A frame that works wide and fails vertical is common, and it is only visible if you look at the vertical.
The habit
A one-line addition that prevents a category of rework.
Write the ratio down before generating anything, alongside where the frame is going and what will sit on it. It takes ten seconds.
Every prompt in the set then carries it, and the whole batch comes back usable rather than nearly usable.
This is the least interesting decision in the process and one of the few that cannot be fixed later without redoing the work.
1:1 · two hours