All guides

AI Video line · stop 02 of 14 · 24 min · members

Keeping a product intact when the model wants to redesign it

Image-to-video will redesign your client's product if you let it. The constraints that stop it.

Free with an account

Sign in to read.

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.

01

The stake

A client will notice one wrong stitch.

Everything else about the frame can be beautiful and the shot is still unusable.

Generated video is forgiving about a great many things. It is not forgiving about the one object in frame that the client manufactures, photographs, and has looked at every day for two years.

They will see it immediately. A logo that has drifted half a centimetre, a bottle whose shoulder has rounded, a fabric weave that changed direction — these read as errors to the person paying, no matter how good the light is.

So product work has a different success criterion from every other kind of generation. The question is not 'is this a good shot'. It is 'did the object survive'. Everything in this guide follows from that.

02

First principle

Start from a real photograph, always.

Text-to-video cannot preserve a product it has never seen.

There is no prompt that describes a specific product precisely enough to reconstruct it. Not a long one, not a technical one. If the model is generating the object from a description, it is designing a plausible object in that category, and plausible is not the same object.

So product work is image-to-video, without exception. The first frame is a real photograph of the real product — ideally one from the client's own shoot, which also means the colour is already approved.

Choose the still carefully

The first frame is doing most of the work, so spend time on it. A cleanly lit, sharply focused, straight-on photograph survives motion far better than a dramatic angled one. Every ambiguity in the still becomes a place where the model invents something during the move.

03

Second principle

Move the camera, not the product.

The object stays still and the world moves around it.

The single most reliable structural decision in product video: the product does not move, the camera does.

A slow push in on a static bottle is a shot the model can do almost perfectly, because the object's geometry stays consistent and only its scale and perspective change. A bottle that rotates is a shot where the model has to invent the side it has never seen — and it will invent it, confidently and wrongly.

This is a real constraint on what you can make, and it is worth accepting rather than fighting. A campaign built from static-product, moving-camera shots looks deliberate and controlled. A campaign built from rotating products looks like a series of near misses.

Where the product genuinely must move — a pour, a fabric fall, a lid opening — treat it as the hardest shot in the job. Budget several attempts, keep the movement small, and be prepared to do it practically instead.

04

The constraint block

Name the specific parts that must not change.

Generic preservation language is much weaker than a named list.

'Preserve the product' is close to useless. The model has no idea which pixels you mean. Name the attributes:

The bottle keeps its exact shape, proportions and colour.
The label stays in the same position, at the same size,
with the same layout. The cap does not change form.
No text on the packaging changes.
No cuts. No additional objects enter frame.

Then add the negative that matters for your specific product. For a garment it is the seam and hem lines. For glass it is the reflection and the liquid level. For anything printed it is the type.

Write this block once per product and reuse it across every shot in the campaign. It becomes part of the product's file, alongside the reference photograph — and consistency across shots is largely a consequence of using genuinely identical constraint text rather than a slightly reworded version each time.

05

Reviewing output

Check the last frame first.

Drift accumulates, so the end of the clip is where failure is visible.

Do not watch the generation from the beginning. Jump to the final frame and compare it against your source photograph, side by side, at full size.

Drift is gradual, which means it is nearly invisible in motion and obvious in a still comparison. A label that has crept two degrees out of alignment reads as fine while playing and as clearly wrong once you freeze it — and the client will freeze it.

A quick review order that catches most problems:

  1. Final frame against the source still. Shape, colour, label position.
  2. Any text on the packaging, read letter by letter.
  3. Scrub for a hidden cut — a sudden change of lighting or angle mid-clip.
  4. Only then, watch it for whether it is a good shot.

Doing this in the other order wastes time falling in love with clips that were never going to pass.

06

When it will not hold

Know the shots to stop attempting.

Some product shots are not currently generatable, and recognising them early is a professional skill.

Consistently hard, in rough order of difficulty:

  • Readable small type — ingredient lists, legal text, fine print. It will not survive.
  • Full rotation — anything that reveals a face of the object not present in the source.
  • Transparent containers with liquid — the refraction changes and the level drifts.
  • Hands interacting with the product — two hard problems multiplied together.
  • Reflective surfaces showing a scene — the reflection is re-invented every frame.

When a brief calls for one of these, say so early and offer the alternative. A static macro with a camera move, or a practical shot composited into a generated environment, will deliver what the brief actually wanted.

The professional value here is not that you can generate anything. It is that you know which shots to generate, which to shoot, and which to redesign — before the deadline rather than after it.