I work on a browser tool for short AI videos, and I want to show the workflow we actually use when a clip has to start from real files, not from a blank prompt. This is a show-and-tell. I would like critique on the method.

The problem we kept hitting with text-only generation is continuity. A product still is perfect in shot one and slightly the wrong shape in shot two. So we built the site around two capabilities, and we try to use them in that order.

First, multi-material input. You can drop in images, video clips, and audio together and give each file a job: this still is the subject, this clip is the camera move, this track is the beat or the voice-over. In the text we tag files as @1 or @2 so the prompt is pointing at a file, not waving at a mood. The example we publish is mundane on purpose: two images, one video, and one audio track becoming an eight-second first-person fruit tea ad.

Second, precise editing. If a take is already close, we do not regenerate the whole clip and hope the camera path comes back. We point at one thing — a product, an outfit, a person, or the background — change only that, and keep the motion and lighting. One documented case is a source video plus a packaging image, with every bag in the clip replaced.

A naming clarification, because this trips people up. Seedance is a ByteDance model family. seedance 3.0 is an independent third-party site. We are not official ByteDance, and we do not claim their endorsement. On the seedance ai page we currently put three of those models in one panel: Seedance 2.5 (up to 30 seconds, up to 50 reference files, 720p on our site, with part-edits), Seedance 2.0 (up to 15 seconds, 4K on our site), and Seedance 2.0 Mini (same 15-second cap, 720p, fastest and lowest cost). The panel shows a credit estimate before you generate, and failed jobs return credits.

Things that still break: faces, printed text, and logos can drift. Some models limit photos of real people. You need rights to every file you upload.

If you try this, I am especially interested in whether the @-tagging of files feels obvious, and whether you would reach for the one-element edit or just hit generate again.