Why does AI video change your product between shots? (and how to stop it)

What is actually happening

When you generate a shot with an AI video tool (Runway, Kling, Veo, Sora…), the model doesn’t “place” your product in the scene. It generates each frame by predicting which pixels fit your prompt and the reference image. The result looks like your product, but it isn’t your product.

This causes four typical problems:

• The logo and type warp. Small lettering is the hardest detail to keep: the model rebuilds it as shapes, not as text.
• The packaging changes proportions. A bottle narrows, a cap changes colour, a box gains or loses an edge.
• The product mutates between shots. Every new shot is a new generation: even if the product came out right in the first, the model reinterprets it from scratch in the second.
• The product gets “invented” as it turns. When the camera shows a side that wasn’t in the reference image, the model imagines it.

AI outputWith the real product

Why a better prompt won’t fix it

Writing “keep the logo exactly the same” helps little, because the problem isn’t the instructions: it’s how the model works. It can reduce errors, but it can’t guarantee the product is identical in every frame. For a brand, “almost the same” isn’t good enough: a misspelt logo is reason enough not to run the commercial.

How I avoid it: my method

1. I start from the real product, from 3 or 4 angles

I ask for sharp photos of the product from 3 or 4 angles (front, side, back and a close-up of the label or logo), plus renders, packaging files or the vector logo if there are any. The more real angles I have, the less the model has to invent.

If the brand has technical drawings or a 3D model of the product, I build a 3D version so the keyframes are even more accurate. When there isn’t one and the project needs it, especially in automotive, I work with a 3D partner. And when the commercial needs a setting that stays consistent from shot to shot, I build a basic 3D environment as a reference.

2. I composite and retouch the keyframes before animating

Every shot starts as a still: the real product placed in the scene with the right light, reflections and shadows, and with the exact type and logo, retouched by hand in Photoshop. This is the step that makes the biggest difference, because the model starts from a product that is already right instead of imagining it.

3. You approve them before animation

You approve the product, the characters and the setting as stills. Fixing a frame is far quicker and cheaper than fixing a video.

4. I animate in short shots

The longer a shot lasts and the more the camera moves, the further the product drifts from the original. That’s why I work in short shots, from the approved keyframes, and generate several takes of each.

5. I discard takes where the product changes

I don’t try to patch a take where the product has warped: I discard it and generate another from the approved keyframe. Only takes where the product is the real one make it to the edit.

6. I edit for continuity

In the edit I check that the product is the same from one shot to the next: colour, size, logo position and reflections.

Can I do it myself with the tools?

Partly. If the product is simple (no small type, simple shapes) and the commercial has few shots, you can improve the result a lot with reference images, short shots and by discarding the takes that fail.

When the product has type, a complex logo, reflective materials (glass, metal, car paint) or has to be seen from several angles, it needs compositing and retouching by hand. It’s the same reason professional product photography has always gone through retouching.