AI product videos where the product stays the same

Why products break in AI video

Text-to-video models have seen thousands of bottles and sneakers, so a prompt produces a plausible product, not yours. Image-to-video does better in the first frame, then drifts as the camera moves: the cap gets taller, the label wraps differently, the stitching changes.

Reference-to-video models such as Kling 3.0 Omni and Veo 3.1 take several photos of the item and treat them as a description of what must appear. That is the difference between a mood video and a product video.

What to upload

Photograph the actual item, not a render, unless the render is exactly what you sell. Three to five images cover most products: the hero angle from your packaging or product page, a straight front or side view, the back, a three-quarter view from slightly above, and a close-up of the logo or label, sharp enough to read.

Shoot on a plain light-gray or white background, with soft, even light and no hard reflections across the logo. Keep the color accurate. If the model sees a warm-tinted photo, it reproduces the tint. Remove price tags, protective film and stickers that are not part of the product.

If the model accepts reference video, a short clip of the product turning slowly on a lazy Susan adds angles that stills miss. Kling 3.0 Omni accepts one reference video of 3 to 10 seconds, according to Kling's guide as of October 2026.

A six-shot list for a brand video

1. Hero reveal: the product on a plinth, camera pushes in slowly, ends on the hero angle. 5 seconds.

2. Detail: a slow macro slide across the logo or texture with shallow depth of field. 3 seconds.

3. In hand: a hand picks the product up and turns it toward the camera. Keep the grip simple. 5 seconds.

4. In use: the product doing its job, such as a pour, a spray, a click or a step. One action only. 5 seconds.

5. In context: the product in a lifestyle setting, such as a bathroom shelf, a desk or a gym bag, with a slow camera drift. 5 seconds.

6. End card plate: the product centered on a clean background with space for your logo and text, added later in an editor. 3 seconds.

That is 26 seconds of footage, which edits down to a 15-second ad or a 6-second bumper. Generate two or three takes of each shot and keep the best.

Writing the prompts

Let the references describe the product and the prompt describe everything else. Do not list the product's features; a second description competes with the photos. Specify one motion, the camera move, the light and the final frame. Ending on an angle you photographed, such as "ends with the front label facing the camera", keeps the model on known ground at the moment viewers look most closely.

Keep the product large in frame. A product that fills a fifth of the frame gets fewer pixels and more approximation.

Pitfalls to check before publishing

Small text. Ingredient lists, legal lines and model numbers are redrawn by every current model. Plan shots where fine print is out of focus or turned away, and add real text in editing.

Proportions in hands. Hands make products look smaller or larger than they are. Compare the in-hand shot with a real photo.

Color shifts under scene lighting. A sunset scene warms a white product. That can be fine for mood, but use neutral light for the shots that sell the color.

Claims. Do not show the product doing anything it cannot do. An AI shot of a jacket shedding heavy rain is an advertising claim like any other.

Variants. A different colorway is a different product. Give it its own reference set instead of asking the model to recolor.

Making the shot list in RefVid

In RefVid you create a product, upload 1 to 5 photos with the hero angle first, and generate each shot from that same set on Kling 3.0 Omni. The shot lengths above match RefVid's 3- and 5-second options, and each clip is listed on the product's page, where a label drifting in one take is easy to spot. Rate the takes you keep so the best ones stand out.