No filming:An AI avatar or clone yourself
    Versely
    Guides

    PIP the talking head over the product

    Film or lipsync the face, I2V the product, then composite. One prompt cannot invent both plates.

    Versely Team••7 min read

    Film or lipsync the face. Image-to-video the product from a locked still. Then composite. Picture-in-picture is two files, not one prompt. Add picture-in-picture to a video (create_ugc_video_overlay) parks a talking-head on a base plate. It will not invent a presenter and a honest pack in a single generate.

    This is the UGC inset procedure: face over product. It is not a reaction-marketplace essay.

    Finish two plates, not one prompt

    If you do not have both files, you are not ready to PiP.

    Base plate (product). The audience came for this. Hands, pour, orbit, texture. Identity of the SKU has to hold.

    Overlay (face). Commentary, reaction, the line. Identity of the person has to hold.

    A generate that tries to put both in one prompt is a different model call with a different identity risk. You get a cousin bottle, a melted jaw, or a presenter whose hands are also the pack. That is not PiP. PiP is a composite of two keepers.

    Before you generate anything:

    1. Lock the product still. Crop to the ship ratio (usually 9:16 for the base).
    2. Write the spoken line. Exact words. This is the overlay’s job.
    3. Decide whether the face is filmed, lipsynced to TTS, or a silent reaction. Silent reaction still needs a real face file.
    4. Decide corner. Bottom-left and bottom-right both fight captions and chrome. You will place after you know where type sits.
    5. Write what must remain visible on the base: label, pour, price overlay. The inset may not cover those.

    Reject a single-file “UGC” take that already has a tiny generated face in the corner. That face is texture. You cannot restyle it, restage it, or recast the VO without regenerating the whole frame.

    The UGC video generator can make either plate. The editor composites them. Keep the jobs split.

    Film or lipsync the face. Do not I2V a stranger over the pack.

    The inset is a performance. It needs a mouth you control or a face you actually shot.

    Door A — film it. Phone, locked-off, 9:16 or 1:1 you will scale down. Same light you can live with against the product plate. Say the line. This is the honest UGC default.

    Door B — lipsync. Still of the real person (or a likeness you have rights to) plus audio.

    Procedure for door B:

    1. Generate or record the line first. Gemini 3.1 Flash TTS is 4 credits, 30 voices, inline tags. Exact words.
    2. Take a still of the face that will appear. Not a text-to-image cousin.
    3. Run a video-to-lipsync row with plate + audio: Kling Lipsync at 7 credits, 4K listed, or VEED Lipsync at 4 credits. Match duration to the line. Do not speed the voice to 1.4× to fit a 2-second inset.
    4. Watch the mouth. If visemes miss, recast the audio or the still. Do not composite a miss.
    5. If you do not need the mouth to match — a reaction, a nod, a silent side-eye — skip lipsync. Film or generate a silent face. Lay the VO on the base or as a bed under the composite later.

    Do not image-to-video a product still and hope a face appears in the corner. Do not text-to-video a “creator holding the bottle” and then also PiP a second face. One face.

    Premade avatars and lipsync create the inset. They do not place it. Placement is this composite.

    Rights: a generated presenter who looks like a real identifiable person is a different legal object than a synthetic nobody. Do not solve that on this page by “using PiP.” PiP does not launder a likeness.

    I2V the product from the locked still

    The base is where the SKU lives. Text-to-video will invent a cousin.

    Procedure:

    1. Attach the approved pack still.
    2. Prompt one verb: unscrews, pours, clicks, rotates. Camera simple. No dialogue in the product prompt.
    3. Pick an image-to-video row that matches the brief. Silent is fine if the voice will live on the face or as a bed. Native audio on the product plate will fight the overlay’s speech unless you mute it later.
    4. Generate. Watch muted. Score the pack: label, proportions, the verb travelling through time.
    5. Trim to the speech length, or slightly longer so the inset can sit over a hold. Trim a video the base, not the composite.
    6. If the label is mush, stop. PiP will not sharpen it. Upscale the base or reshoot the still.

    Duration: the overlay will be scaled down, but it still has a clock. A 6-second face over a 3-second pour will hang on a freeze or loop depending on the tool. Match lengths on purpose. Pad the base with a hold on the pack rather than looping a pour.

    If the product is a screen recording or gameplay, that file is already the base. You still need a real overlay file. The parameter name slideshow_video_url on this tool accepts any base video, not only slideshows.

    Composite, then caption around the inset

    create_ugc_video_overlay wants a base URL, an overlay URL, a required position (a corner, not a free pixel), and a size/scale.

    Procedure:

    1. Confirm both keepers. Do not composite maybes.
    2. Set position. Default temptation is bottom-right. On 9:16 that is also the caption band and often the CTA. Prefer a corner that leaves the pack label and the future caption row clear. You will verify on a phone.
    3. Set overlay size small enough that the product remains the plate. A talking head that covers the pour is not PiP. It is a face with a sliver of B-roll.
    4. remove_black_background only if the overlay was actually rendered on black. A different-coloured backdrop needs a real background-remove first.
    5. Render. Watch muted. Does the pack still read. Does the face sit off the label.
    6. Listen. If both files had audio, you now have two tracks or a replace. Decide: overlay mic, TTS bed, or mute the base. Replace video audio on the composite if the mix is a mess.
    7. Burn captions on the composite. Add captions to video. Place type off the inset and off safe area chrome. A karaoke accent on the mouth you already composited is two mouths.

    If you needed several base clips plus an inset, compose is the fuller tool. This page is one product plate plus one face.

    Do not caption the overlay and the base separately and then composite. Type in the inset scales down into junk. Type on the base gets covered by the inset. Caption last, on the joined picture.

    Corner-mount the keeper face over the keeper pack in Versely. Bring two URLs. Do not bring one prompt for both.

    FAQ

    Can I prompt a model to generate PiP in-camera?

    You can. You will not get two restyleable plates. PiP is a composite of files you already have.

    Does the face have to lipsync?

    Only when the mouth is the point. A voice-over bed over product B-roll with a small silent reaction still works. Lipsync when visemes must match the line.

    Where should the inset sit?

    In a named corner that does not cover the SKU or the caption row. Check on a phone against current platform chrome. There is no permanent pixel spec.