Kombucha Brand: Keep the Label Readable in AI Video With a First Frame

How do you keep a kombucha label readable in an AI product video? Never ask the model to invent it. Give it your real label photo as the first frame (image_url) and describe only the motion. On Sume, gemini-omni-flash-1.1 renders 8 seconds at 720p for $1.00, with native audio always on.

Small printed text is where generated video fails first. A brand name that morphs mid-clip looks like a counterfeit, so the plan is to protect the label rather than hope for it.

Photograph the can, then animate around it

Take a flat, well-lit photo of the bottle or can with the label square to the camera. Send it as image_url in a Video Router request for gemini-omni-flash-1.1, which Sume routes as image-to-video. Write a prompt about the environment and the motion: condensation beading, a slow orbit, ice falling beside it.

Keep the move gentle. A slow push-in or a small sideways drift keeps the label facing the lens. A fast spin forces the model to invent the back of the pack, and invented printing is exactly the failure you are avoiding.

Why Omni is a reasonable pick

gemini-omni-flash-1.1 accepts 3 to 10 seconds, resolutions of 360p, 720p, 1080p and 4K, and aspect ratios of 16:9 or 9:16 only. It accepts up to 10 reference images and up to 3 reference videos of 3 seconds or less, in reference mode, and it has an edit mode that works on an existing clip through video_url. A request cannot mix video_url, image_url and reference lists. Audio is native and always on, so fizz and a pour arrive in the clip.

Its list prices are $0.03 per second at 360p, $0.10 at 720p, $0.15 at 1080p and $0.30 at 4K. Sume bills list times 1.25, so 8 seconds at 720p is 8 x $0.125 = $1.00, matching the golden test.

Check the label like a customer

Scrub through the clip frame by frame and read the label on each. If the name holds for the whole clip, you are done. If it softens in the middle, shorten the duration: a 5-second clip gives the label less time to drift than an 8-second one.

Draft at 360p first, since 8 seconds there costs only $0.30. Move to 720p once the motion is right, and use 1080p only for a hero shot where the label is a feature.

What generation will not fix

If the label must be pixel exact, use generation for the background and the real photo for the pack. A clean cutout of the can from your photo can be laid over a generated scene in an editor, so the printing never passes through a model at all.

Cutting out a real photo is a separate step. The honest result is a clip where the pack looks like your pack, because it started as your pack.

Flavour variants without reshooting

A brand with five flavours does not need five photo shoots for video. Photograph each can once, then reuse the same prompt template and swap the first frame. Five 8-second 720p clips cost 5 x $1.00 = $5.00, and the motion looks consistent across the range.

Omni can also edit an existing clip through video_url, which suits a hand-filmed pour that needs a different background. Reference videos must be 3 seconds or shorter, so cut them first with Video trim at $0.02 each.

Remember that Omni takes only 16:9 or 9:16. Vertical suits a feed, horizontal suits a retailer page, so plan both frames from the start.

Finish with captions if the clip will play muted in a feed, using Video captions at $0.20. A short line such as the flavour name helps a viewer who cannot read the can at phone size, and it gives the brand name a second chance to appear exactly as spelled.