Upload a clip and a mask, and an AI model reconstructs what was behind the object you erased, frame by frame, with no rotoscoping and no timeline editor. Built for the passerby in the background, the cable in a product demo, or a watermark you never licensed.
A stranger walks through the back of an otherwise perfect shot. A microphone stand sits in frame during a product demo. A watermark from a stock clip sits in the corner of the footage you now need clean. Rotoscoping that out by hand, frame by frame, used to be the only fix, and it ate hours nobody had. A model built for this job removes the object and reconstructs what was behind it, using the pixels around the gap, in minutes instead of an afternoon.
Manual object removal in video is rotoscoping: tracing the unwanted element by hand on every frame it appears in, then painting or cloning over it, then checking the seam does not flicker. It is a real, teachable editing skill, and it is also the kind of repetitive, precise work that eats a disproportionate share of a post-production schedule for something the audience was never meant to notice.
Picasso IA runs bria/video-erase-object, a model built for exactly this: you supply the clip and a mask marking what to remove, and it fills the erased region using the surrounding visual context, frame by frame, so the result holds together from the first frame to the last instead of drifting or flickering the way a rough manual patch can. It sits in the video-editing category of the catalog alongside models for background removal, upscaling and text-guided restyling, so the same account that erases an object can also clean up the footage around it.
The model takes two inputs: your video and a mask video that marks, in white on black, exactly which pixels to remove and on which frames. It reconstructs the masked region from the context around it, keeping movement and lighting consistent across the clip, and by default it keeps your original audio track untouched, since removing something from the picture rarely means removing it from the soundtrack too.
Output goes out as MP4, WebM, MOV or MKV, with a choice of codec including H.264, H.265, VP9 and ProRes, so the file that comes back matches whatever you are editing in next rather than forcing a conversion step. One real constraint shapes how you use it: the model processes up to five seconds of footage per run, and an auto-trim option will cut a longer upload down to its first five seconds rather than reject it outright, which is worth knowing before you upload a thirty-second take expecting the whole thing back.
The mask is doing the real work, and what counts as a good mask changes with the target. This is the honest version of what to expect for the removals people actually ask for.
| What You're Removing | What to Mask | Worth Knowing |
|---|---|---|
| Passerby in the background | Their outline on every frame they appear in | A person moving fast needs more mask frames than one standing still |
| Logo or watermark | The exact printed shape, frame by frame | Run video upscaling after if the patch looks soft |
| Mic or cable in a product demo | Just the object, not the hand or surface near it | A tight mask keeps nearby texture from being erased too |
| Moving vehicle in a street shot | The vehicle's silhouette through its whole path | Fast motion can leave a mild blur where it passed |
| Notification bar in a screen recording | The fixed rectangle, held for the whole clip | The easiest case, since the mask barely moves |
This is the loop from a raw clip to a clean one. It assumes nothing beyond the clip itself and a basic video editor to draw the mask in, and it takes longer to read than to actually run.
The mask is the one input you fully control, and most disappointing results trace back to it rather than to the model. A few habits make the difference between a seamless erase and a visible smudge.
Draw the mask a few pixels wider than the object rather than tracing its outline exactly. The fill reconstructs from the pixels just outside the mask edge, and a mask that hugs the object too closely tends to leave a faint ghost of it behind.
An honest tool page says where the tool stops working. For video object removal there are real limits worth planning around before you rely on it for something client-facing.
The five-second processing window is the most immediate one: anything longer needs to be split, processed in pieces and merged back together, which is an extra step compared to a single pass. There is no automatic object detection or tracking either, so the model does not find the thing you want gone and draw the mask for you; that part stays manual, and it is the slowest part of the whole workflow. Heavy occlusion, where the removed object crosses in front of other moving elements for several frames, tends to produce a softer, less certain fill than a clean pass over a static background. And this model erases and reconstructs; it does not add or replace anything, so putting a different object in the same spot is a separate job for a text-guided editor in the same catalog category, not this one.
For a single still frame rather than a moving clip, an image-based object removal workflow is the simpler and cheaper tool, and it is worth knowing the two are different products built for different inputs. For a full studio comparison of editing depth across tools, the Picasso IA vs Runway breakdown covers where a 488-model catalog and a single specialized platform each pull ahead.
Real estate agents use this to pull a stray car or a neighbor's bin out of a listing walkthrough before it goes live, which pairs naturally with the broader workflow on the real estate video tours page. Product teams use it to clean up a demo clip shot with a visible mic or rig rather than reshooting the whole take, a use case the product demo videos page covers from the filming side. Social and marketing editors reach for it on footage licensed from a stock library that still carries somebody else's watermark in the corner.
It replaces frame-by-frame rotoscoping, cloning and seam-checking with one automated pass: you mark the region once as a mask, and the model reconstructs it consistently across every frame instead of you painting each one by hand. The editing skill still matters for building a good mask, but the tedious, repetitive part of the job moves to the model, which is usually the part nobody enjoyed doing anyway.
You need some way to draw a white shape over the unwanted region on a black background, matched frame by frame to where it sits, which most basic video editors can do with a shape or rotobrush tool. It takes more editing skill than typing a text prompt, but far less than manually painting over the object on every frame, since the mask only needs to mark the area, not reconstruct anything.
Video Erase Object processes up to five seconds per run, and an auto-trim setting will cut anything longer down to its first five seconds automatically rather than fail the job outright. For a longer video, split it into five-second segments with a trimming model, run each one through separately, and merge the results back together afterward.
Yes. The video-editing category in the 488-model catalog covers background removal, upscaling, text-guided restyling and more alongside object erasure, including models from Runway and other specialist publishers, so an account built around erasing objects can also handle the rest of a clip's cleanup. The Picasso IA vs Runway page compares the two approaches directly if you are choosing between a catalog and a single dedicated platform.
Not by default. The model preserves your original audio track unless you turn that off, so removing a visual element leaves the soundtrack exactly as it was. That is usually correct, with one exception worth catching before you generate: if the object you erased was itself making the sound, such as a speaker or a device, the audio will still play as if it were there.
Generally yes, since a printed logo or watermark tends to have a fixed, well-defined shape that is easy to mask precisely. The main variable is what sits behind it: a plain background reconstructs cleanly, while a logo overlapping a busy, detailed scene can leave a faint softness where the patch sits, which an upscaling pass usually tightens up.
Video Erase Object takes an explicit mask and removes exactly what you marked, which gives precise, predictable control over one specific region. Text-guided editors elsewhere in the catalog work from a written description instead, and they are better suited to restyling a whole scene or swapping one object for another than to a clean, targeted removal, since they are not working from a pixel-accurate mask.
Sometimes, and it correlates strongly with the mask and the background rather than being random. A loose mask, heavy occlusion behind the removed object, or fast motion through the region can all leave a softer patch than a static shot over a plain background would. Running the clip through an upscaling model afterward tightens up most of the mild cases.
Generations are paid in credits, and the exact cost depends on the model and clip length, so any specific number written here would be wrong within a few months. New accounts start with free credits, which is enough to test the workflow on a real clip before committing to anything, and current plans live on the pricing page, which is the only place worth trusting for numbers.
Yes, and those are two of the most common real uses. A listing walkthrough with a stray car, bin or reflection in frame, or a product demo shot with a visible mic or rig, are both exactly the kind of single, well-defined object this model handles well, without reshooting the take.
A cluttered shot is not a wasted one. Open the Toolkit, build a mask around what does not belong, and see what the frame looks like without it.
