Gallery and auction video without touching the art
Generatively altering a work is disqualifying, so the craft is camera move, light and context. A single-object push-in and provenance language that holds.
Run a painting through an image-to-video model and watch the brushwork. The impasto reorganizes itself. A signature becomes a smear that resembles a signature. Craquelure crawls. The model is not moving your painting through space; it is generating a new painting, many times a second, that looks broadly like yours.
For almost every other product category this is a non-issue: a slightly reinterpreted sneaker is still a sneaker. Here it is disqualifying. The object's surface is the value, condition language is contractual, and a buyer who bids on the strength of your video and receives something whose surface does not match has a complaint that goes well past marketing.
So the constraint defines the craft. Everything interesting comes from where you put the camera, how you light the thing, and what you build around it.
The push-in happens in the edit, not in the room
The signature move for a single lot is a slow push-in: start on the full work, drift in over eight to twelve seconds, land on a detail. It mimics how a person actually looks at a painting, and it holds attention without narration.
Do not make it with a camera. A slider, gimbal or tripod being repositioned near a consigned work is risk for no benefit, and a physical move adds focus drift and perspective shift that make the piece harder to read.
Make it in post instead:
- Photograph the work at three to four times your delivery resolution. For a 1080-wide vertical delivery that is a 3,240 to 4,320 pixel capture, which lets you push in threefold and still sample native pixels rather than interpolate.
- Lock the camera off. Tripod, remote release, mirror-up, no one touching anything.
- Animate the crop in the edit. The push-in is a scale-and-position ramp on a still. No generation, no re-synthesis, no model deciding what a brushstroke should look like. Every pixel that reaches the viewer came off your sensor.
- Ease the ramp at both ends. A linear zoom reads mechanical.
- Output at 25 fps, the default, and keep it consistent across a sale so lots cut into a preview reel without a frame-rate mismatch.
Turning a slideshow into a video is the mechanism for the still-to-motion assembly, and upscaling image resolution covers the case where an archive image is the only capture you will ever get. Image-to-video is the right tool for a product on a turntable and the wrong tool for a Rothko, and knowing which is which is most of the professional judgment in this vertical.
Lighting a single object so the video matches the condition report
The video and the condition report have to describe the same object. When they do not, the video is the thing the buyer remembers.
- Cross-polarize. A polarizing filter on the lens and polarizing film on the lights kills specular glare on varnish, glazing and glass. That is the difference between a usable frame and one with a window reflected in it.
- Shoot the raking-light pass as a separate, labeled clip. Light skimming across the surface at a shallow angle reveals impasto, canvas weave, cupping and losses. It is enormously informative and it is not what the work looks like on a wall, so it goes in its own segment with a caption saying so. Presenting a raking-light frame as the general appearance is a condition misrepresentation made with entirely real photons.
- Use consistent, high-CRI continuous light across the whole sale. Lots lit differently look like they came from different collections, and colour that shifts between lots undermines every colour statement you make.
- Put a reference target in a frame you do not publish, set white balance from it, and log the setup so the next session matches.
- Put "colour is indicative" on screen, every time, and point to the condition report. No display chain reproduces a painting faithfully and everyone in the trade knows it.
For three-dimensional work, same discipline with a turntable instead of a crop ramp. For anything under glass, shoot with the room black and the lights well off-axis.
Provenance and condition language that survives a dispute
This is where the video stops being marketing and becomes something closer to a document.
The governing rule: the video may not describe the lot more confidently than the catalogue does. Catalogue vocabulary is not decorative. "By," "attributed to," "studio of," "circle of," "after" and "in the manner of" are a graded scale with meaning, and a voiceover that flattens "attributed to" into "by" has made a different representation than the one your specialist signed off on.
Practical rules that follow:
- Read the catalogue line verbatim, on screen and in narration. Not a friendlier paraphrase. If the line is awkward, fix it in the catalogue and use the fixed version in both places.
- Never let a language model rewrite a catalogue entry for a caption. Paraphrase is exactly the failure mode: it produces text that reads better and asserts more.
- Provenance is a list, and it stays a list. Named collections, dates, and gaps shown as gaps. "Private collection, Europe" is a real entry; inventing continuity to tidy it is not.
- State that the video is not a condition report and does not replace inspection, in the description on every platform.
- Know what your written descriptions create. In some jurisdictions a written statement of authorship supplied by an art merchant to a buyer creates an express warranty, and your published conditions of sale govern the rest. Specialists and counsel own that language; marketing transports it unchanged.
Two copyright facts that catch galleries out
Owning the object is not owning the copyright. Transfer of a physical work does not transfer the rights in it, so for most work by living artists and many estates, copyright sits with the artist or the estate rather than with the consignor or with you.
And the exemption that lets you hang a work on your wall and let the public look at it is limited to viewers present at the place where the copy is located. It does not extend to a video posted to Instagram, embedded on a lot page or emailed to a bidder list. That is a different act.
The consequence is administrative rather than dramatic: your consignment and artist agreements should grant the reproduction rights you actually need, in the media you actually use, and someone should check that they do before you build a video programme on top of them. Museums face a near-identical problem and usually solve it with clearance built into the loan agreement — AI video for museums and cultural institutions covers that side, and the museum guide has the broader picture.
What you may generate
Plenty, once the work itself is off the table.
| Element | Approach |
|---|---|
| The work, any part of it, any detail | Photographed. Never generated, never model-altered |
| Signature, verso, labels, stamps, frame detail | Photographed. Generating one of these fabricates evidence |
| Sale identity cards, lot number, estimate, catalogue line | Typed overlays in the editor |
| Artist-context b-roll: the city, the period, the studio type | Generated, visually distinct from the work, and labeled |
| Narration for context pieces | Voiceover |
| Translated captions for international bidders | Generated from your own script |
| Installation and viewing-room footage | Filmed |
The context b-roll is where generation genuinely helps, and it needs a hard visual separation from the lot photography so no viewer could mistake one for the other: a different aspect treatment, a grade that reads as archival, or a persistent corner label. Do it the same way every time and disclose it — synthetic media disclosure covers the labeling conventions.
Do not generate a person who could read as the artist, a period photograph that could be mistaken for documentation, or an exhibition that did not happen. Where provenance is the product, fabricated context is worse than no context.
Build the sale as one timeline structure: title card, lot card, push-in, raking-light segment, catalogue line, close. Save it as a reusable draft and swap the photography and the typed layer per lot, and a hundred-lot sale becomes data entry rather than a hundred edits. The 480p preview pass is free with a short per-user cooldown, which is the right place to catch a wrong lot number before the charged final export. The gallery guide has the platform-level view, and saving an edit as a reusable draft is the mechanism.
FAQ
Can we ever run a work through a generative model, even for a thumbnail?
No, and thumbnails are where someone will try, because a thumbnail feels disposable. It is not: it is the frame that appears in search, in a share preview and in a bidder's email, at a size where nobody looks closely enough to notice the surface is wrong. Crop your own photograph — generating a video thumbnail from your own footage is the correct path.
How long should a lot video be?
Twelve to twenty seconds for the social cut: one push-in and the catalogue line. Forty-five to ninety seconds for the lot page, adding the raking-light pass, a detail or two, and the provenance list held long enough to read. The specialist piece explaining why a lot matters can run three to five minutes and belongs on the sale page. All three come from one capture session; only the assembly differs.
Does upscaling an old archive image count as altering the work?
It depends on what you upscale and how you present it. Enlarging a low-resolution photograph so it holds up at delivery size is a reproduction decision. A generative upscaler invents plausible surface detail that is not in the photograph and may not be in the painting, and presenting that as documentation is the thing this post is about avoiding. If a generative pass touched a lot image, either do not use it or label it as an enhanced archival reproduction rather than as a record of the work.