Change a person's outfit in a video: Omni edit or H3 Max Recast

To change a person's outfit in an existing video on Sume, use a Gemini Omni Flash 1.1 edit when you can describe the clothes in words, and use H3 Max Recast when you need a specific person from a photo. Omni costs $1.25 for 10 seconds at 720p. Recast costs $3.75 for 10 seconds at 768p and $5.625 at 1080p.

What each route actually changes

The two are not interchangeable. An Omni edit takes video_url and a text prompt only; it cannot take a garment photo, because video_url cannot be combined with reference fields. Recast takes a source video_url and one to four reference_image_urls, one photo per person, and its job is to replace the people in the clip with those photos. The outfit comes from whatever the photo shows.

Rates for a 10-second clip

Omni output follows the source length for edits. Recast's duration equals the source length, 5 to 30 seconds.

When to pick the text edit

Choose Omni when the brief is a category of clothing (a navy blazer over a white shirt, a winter coat) and the face and body in the clip should stay as they are. It is the cheaper route and it leaves the person alone. The prompt should name the garment, its color and what stays unchanged: "Change the jacket to a navy blazer. Keep the face, hair, movement and background the same."

When to pick Recast

Choose Recast when the person must change too, for instance a model wearing a specific garment, shot once and shown on several people. Supply a clear photo that includes the outfit. For a clip with two people, send two photos in the order described in the 1-to-4 photo guide.

Recast also lets you test cheaply: render at 768p first, then repeat at 1080p once you like the result. The upgrade cost is $0.1875 per second more.

Prompt wording that holds up

For either route, describe the garment by cut, fabric and color in one clause, and keep the preservation clause short and literal. Long lists of things to keep tend to compete with the thing you want changed. Ask for one garment per pass; if the shirt and the shoes both need to change, do the shirt, check it, then do the shoes. Costs add up, but a failed two-change pass costs the same as two clean one-change passes.

On Recast the prompt is optional. Leave it empty on the first run and let the photo drive the result, then add a prompt only to fix what you see.

Verify the garment, then the hands

Outfits fail at the edges: sleeves that change length during a gesture, a collar that disappears when the head turns, logos that morph. Extract stills at three or four timestamps with video-frames and compare them, rather than judging from the first second.

Neither route offers a way to guarantee a brand-exact garment, and nothing in the docs promises it. If the garment has to match a catalog item, treat the output as a mockup.