Turn the keeper into stills

Extract frames from the file you will ship. Turn a video into still photos is extract_video_frames. A thumbnail is one poster frame. Stills are a contact sheet: several reusable image URLs from the keeper, for PDP, for a carousel, for an I2V parent, for legal. Do not text-to-image a pack that already moved correctly in the cut.

This is video → photos from the signed file. It is not stills → slideshow, and it is not “thumbnail from the keeper” as a single YouTube poster.

Stills from the file you will ship

Identity has to match. A prettier stranger is a different ad. A cousin bottle is a listing problem.

Procedure:

1. Lock the video. Extracting from a take you will recut gives you stills of a film that will not exist.
2. Use the shipping file. If you merged hook + demo + CTA, extract from the merge. If you upscaled, extract from the child. If you captioned, decide whether you want type in the still. Burns are pixels; they will be in the JPEG. For a clean pack still, extract from the uncaptioned parent.
3. Set start_time / end_time if you only want the verb, not the whole runtime.
4. Set frame_count or fps. Five evenly spaced frames is a contact sheet. fps 1 over a 6-second pour is six frames of the same verb. Pick.
5. output_format for the file type you actually need.
6. Review at 100%. Soft, motion-blurred, or warped-label frames are not keepers. Delete them. Do not upscale mush and call it print.

The tool returns image URLs. They are assets. Save the ones that pass. The rest are a contact sheet you needed in order to choose.

Resolution is capped by the source. A 720p generate yields a 720p still. Grok Imagine Video is 7 credits, 720p, native audio. Extracting a “4K still” from it does not create 4K. Pair with upscale image resolution if the still must print or survive a later I2V. Upscaling enhances existing detail; it does not invent a sharp flavour name that was already mush.

Do not extract from a 360p Omni draft you made on Google’s surface and treat it as the campaign still. That draft is a working print. The hero file is the parent.

Several frames, not one thumbnail

Generate a video thumbnail is the same extractor with frame_count 1 at a timestamp you choose, aimed at a slot: Shorts cover, Reels preview, YouTube 16:9 poster. This page is the many-frame job.

Why several:

• Contact sheet. Pick the legal pack angle after the fact. The generate’s “hero” moment is often not the sharpest label.
• Carousel. Three stills from one pour: closed, mid, empty. Sequence them in a slideshow later. Do not T2V a carousel.
• PDP / retail. Marketplace wants stills. The video already proved the SKU if you extract from the keeper rather than generating a cousin.
• QC. Pause-worthy frames expose warp, extra fingers, a melted logo you missed at 1× playback.
• I2V parent. A frame where the pack is sharp and the angle matches the next verb is a better contract than a random first frame.

Even spacing is a start, not a strategy. If you know the verb is at 0:04, extract a burst around 4 seconds rather than five frames across a 15-second talking-head. frame_count without a tight window wastes the job on smiles.

Motion blur: a pour at 1/24 of a generate is often a smear. The usable still might be the frame before the verb. That is still the product. Prefer it over a heroic blur.

Faces: a still of a talking head mid-word is unflattering and, if you will use it as I2V, a bad mouth contract. Pick a closed-mouth frame if the next job is lipsync or a silent hold.

Cost is per extraction job. estimate_cost if you are dumping fps 4 from a 10-minute webinar. That is a different brief. This page is a UGC keeper.

Upscale a still, not the whole clip, when the still is the job

If the output is photos, do not upscale the video to 4K first “to get better stills” unless you also needed the 4K video. You will pay for every frame’s upscale and then throw most of them away. Extract, pick one or two, upscale those.

Zoom into type, edges, hair. Compression from the generate will sharpen into crunch. If the label is already illegal, upscale will not typeset it. Go back to the product still you should have I2V’d from, or photograph the SKU.

Using an extracted still as a new generate parent:

• generate_image_from_image for a cleanup (background, colour) — review edges.
• generate_video_from_image / I2V only if the still still is the SKU. A warped extract teaches warp.

Do not extract a captioned frame and ask an image model to “remove the words.” You will smudge the pack. Extract from the uncaptioned file.

A slideshow is the opposite pipeline: stills you already own, sequenced, converted to MP4. If you extracted stills in order to make a carousel video, that is turn-slideshow-into-video, not a new T2V.

Use a still as I2V parent only if identity holds

Pause. Read the label. If a shopper cannot, you do not have a pack shot. You have a frame of motion.

I2V will treat that still as frame one. A 9:16 extract from a 9:16 keeper keeps the window. A 16:9 extract used for a TikTok I2V will generate landscape again. Crop the still to the ship ratio before you attach it.

Rights: the still inherits the video’s rights. A licensed clip does not become a stock photo you sell because you extracted it. A generated clip may still need Article 50 labelling (2 August 2026) when shown as a photoreal image at first exposure. The still is not “not AI” because it is a JPEG.

YouTube custom thumbnails are a slot. Extract a 16:9 frame for a 16:9 poster. Extracting a 9:16 hook frame and uploading it as the watch-page image is the wrong object.

Turn the keeper into stills in Versely. Extract from the file you will ship. Pick the sharp frames. Do not T2V a pack you already have.

FAQ

How many frames can I pull at once?

frame_count or fps, bounded by the time range you set. Pull what you will review. A dump is not a contact sheet until someone looks.

Can I feed an extracted frame into I2V?

Yes, if identity holds. The URL is a normal image input. Crop to the ship ratio first. Soft type will not sharpen in motion.

Will the still be higher resolution than the video?

No. It is capped by the source. Upscale the chosen still if you need it larger. Do not expect 4K from a 720p parent.