Eyewear OEM: the frame still cannot drift

Identity.

For an OEM, the frame is the SKU: temple curve, rim thickness, hinge, the acetate colour that does not exist in "tortoiseshell-ish." A generate that almost looks like last season's 54mm is a different product. Motion is allowed. Morphing the object is not.

The catalog row built for that hold is Seedance 2.5 image-to-video: animate a single still into a native clip at up to 720p, 4s–30s, optional second image as the closing frame, so you are not stitching short takes that drift. requires_image is true. Pick it inside image-to-video: your upload is frame one. Labels, edges, and layout survive because they were already in the file.

The buyer-facing cousin is AI video for eyewear brands: people need to picture the frame on a face. OEM marketing has a stricter prior. The frame in the plate has to stay the frame you tool.

The frame is the identity

Do not start from a paragraph. "Premium titanium optical, subtle Japanese geometry" is how you get a handsome pair that is not yours. Shoot or retouch one still of the actual SKU — front, three-quarter, or on a form — and make that frame one.

Image-to-video is explicit: the still is the opening frame, not a style hint. Prompt only the motion. A slow turntable, a light sweep across the lens, a gentle push. Do not prompt a new silhouette.

The eyewear brands page is right that a white-background pack shot does not answer "will this suit my face." That is a different clip: many real faces, one real frame. OEM sell-in, a line sheet, a factory tour teaser — those can be the object only. Either way the object cannot wander.

If you need a last composition (front hold to three-quarter hold), add the second still as a last frame. That is interpolation with two approved photos, not a hope that a 10-second prompt remembers the endpiece.

One still, not a new face every second

Short multi-clip workflows are how identity dies. Clip one is the SKU. Clip two, generated from a new prompt, is a cousin. Stitch them and the temple width breathes. Seedance 2.5 image-to-video's own description is the warning: one still into continuous motion without the drift or stitching of shorter multi-clip workflows.

Stay on one take. Durations run every second from 4 through 30. Resolution tops out at 720p (480p or 720p). Do not prompt 4K. You will not get it on this row. Aspects include 9:16 — generate that when the destination is a reel, do not crop a wide pack shot and shave the rims.

Native audio is on. If you do not want a invented score over a silent turntable, say so in the prompt, or pick a silent row. Do not discover a soundtrack at export.

Drift is a different SKU

Pause at 0:00, the midpoint, and the last frame. If the front at the end would not pass a line-sheet check against the photo you uploaded, the take is unusable. Tightening the prompt with more adjectives will not put the hinge back. Go back to the still. Shorten the duration. Reduce the motion. A 4-second light sweep on a locked plate is a finished clip. A 30-second "cinematic orbit" is how the model invents a second pair.

Do not use a talking founder as a substitute for the object. A founder can hold the frame. The frame still has to be the frame. If the only way you can keep the silhouette is to cut to the photo, cut. Identity is the point of the generate, not a vibe you add after.

FAQ

Why not Veo the glasses from text?

Because text will not pin the SKU. Veo is a scene row. Eyewear OEM is a product row. Start from the still of the pair you actually make.

Can I put the frame on a generated face?

Only if you accept that the face is marketing and the frame must still match the still. The eyewear brands angle is many real faces wearing one real frame. A generated on-face try-on that changes the rim is a different SKU. Prefer a real face, or keep the clip on the object.

Where do I run Seedance 2.5 image-to-video?

Inside image-to-video. Pick the Seedance 2.5 image-to-video row in the catalog. The still you upload is frame one.

How long should the take be?

As short as the motion needs. The row will let you go to 30 seconds. Identity usually fails late. Start at 4s–6s on the locked still. Extend only when the silhouette still matches the photo.