You remove unwanted objects with AI in-paint and cleanup: select or describe the object, and the model regenerates the pixels behind it across every frame. For isolated objects, use in-paint directly; if the problem is the whole environment, use a swap instead; for objects tangled in complex motion, regenerate the shot.
Start by classifying what you're removing, because that decides the tool. A single unwanted object — a sign, a bystander, a cable, a logo — is an in-paint job. A cluttered or wrong environment around a subject you want to keep is a swap job. An object that crosses or occludes your subject through heavy motion is usually a regeneration job.
In-paint and cleanup for isolated objects. In-paint models let you work directly on existing footage: mark or describe the object, and the model removes it and reconstructs the background behind it frame by frame, keeping the reconstruction temporally consistent so the patched region doesn't flicker. Google Omni Flash ships in-paint and cleanup natively — insert or remove elements from footage you already have — and this capability has become a baseline expectation for Omni-class models rather than a premium add-on. invideo is an agentic video creation tool with the current models available, so you can describe the removal in plain language and the invideo agent routes the clip to a model that supports in-paint rather than you picking one per task.
Swap when the environment is the problem. If you'd end up in-painting half the frame, replace the environment instead of erasing pieces of it. Omni's swap feature replaces backgrounds, environments, or clothing around a subject while preserving the original subject's roto and edges — you keep the person exactly as filmed and regenerate everything around them. This is the faster path when the 'unwanted object' is really an unwanted location.
Regenerate when motion defeats in-paint. When the object occludes your subject or moves through complex action, patching it frame by frame degrades: either mask it manually in a desktop editor with motion tracking, or regenerate the shot entirely with a reference-to-video model such as Seedance 2.0, which carries your subject's appearance from a clean reference frame into a new take without the object.
Check the patched region before delivery. AI-reconstructed pixels inherit the model's texture quality, so inspect the cleanup at your delivery resolution. On Omni Flash, output defaults to 720p with a 1080p upscale at no cost, while 4K costs the equivalent of a full generation — budget accordingly if the cleaned clip is going to a large screen. Generation lengths run 4, 6, 8, and 10 seconds, so process longer footage in segments and cut them back together in your film's aspect ratio.
Watch some of these to see what works for you:
the current state of the textures that Google is offering, I'm not so sure if they're ready for prime time cinema yet.
— invideo's creative team, from testing 30+ Omni Flash outputs