AI video quality checklist for architecture: what to review before you publish

An AI-generated architecture film can look finished and still be wrong in a way a client or a juror will notice: a mullion that bends in the third scene, a stone wall that turns to stucco mid-shot, a figure walking through a glass balustrade. None of that shows up in a thumbnail. This article is a pre-publishing review checklist for AI video of buildings and interiors: what to check, scene by scene, from geometry and materials to music licensing and disclosure wording, with a printable table and a short procedure for what to do when a scene fails.

Key takeaways

• Review AI architecture video at full size, paused, scene by scene; the problems that matter are small and vanish at thumbnail scale.
• Geometry comes first: verticals, window mullions, stair runs and railings are where image-to-video models most often drift.
• Lighting and materials have to stay consistent across every scene, not just within one clip, because viewers read the film as a single visit.
• Any people in the film must be fictional and plausible; your source photos should never show real clients or staff being animated.
• When a scene fails, change the source image before you change anything else; the image is the input the model actually works from.
• Confirm music rights, platform specs and your disclosure line before you export, not after the post is live.

Review in three passes

A still rendering is checked once. A 30-second AI film has six or more scenes, each generated separately from a single image, and each can introduce its own error. Review in three passes: watch the film once at normal speed and note anything that pulls your eye; step through each scene paused at the first frame, the midpoint and the last frame; then check everything that isn't picture. The sections below follow that order.

Geometry: verticals, mullions, stair runs

Buildings are mostly straight lines, which makes them unforgiving. Pause at three points in each scene, full screen.

Verticals. Column edges, door jambs and wall corners should stay parallel to the frame edge, or converge consistently if the camera is tilting. A vertical that bows outward and then returns is an artifact. The usual cause is a lateral camera move; see camera moves ranked by how well geometry survives for why static shots and push-ins hold lines better than pans and trucks.

Mullions and curtain wall grids. Count the bays at the start and end of the scene. The count should match and the spacing should stay even; a grid that gains a bay as the camera moves is a common failure on long facades.

Stair runs and railings. Count treads, check that the handrail follows the slope at a constant height, and check that balusters don't multiply or merge. Open-riser stairs and glass balustrades fail most often because the model has to maintain what shows through them.

Roof lines and reflections. A roof edge should stay straight; a kink that appears mid-scene reads as a construction defect to anyone who builds. A reflection showing a different sky or an invented building fails the scene.

Materials and lighting: texture drift and consistency across scenes

Texture drift is the gradual change of a surface during a clip: board-formed concrete that smooths out, cedar siding that loses its board lines, terrazzo that turns into plain polished concrete. It happens most on large, evenly lit surfaces and on long clips. Check each major material at the first and last frame: same material, same pattern scale, coursing still lined up. A brick wall whose courses change height is wrong even if it still reads as brick.

Also watch for invented finishes: gloss added to a matte floor, a painted steel column turned stainless. These are design changes, and in front of a client or a jury they misrepresent the project. A sharp photograph of the actual finish gives the model less room to improvise; choosing a six-view set that covers a project explains why cropped or compressed images produce more material errors.

Lighting needs the same cross-scene check, because the viewer experiences six separately generated scenes as a single visit. Late afternoon in scene one and overcast noon in scene two reads as two different days, fine for a construction-progress film and wrong for a completed-project film. If the sun moves 90 degrees between two shots of the same elevation, pick different source images. Warm tungsten in one room and cool daylight in the next is plausible if the spaces differ, not for two views of the same room. Flicker within a scene, where brightness pulses as the camera moves, is an artifact and fails the scene.

People, text and signage

A small cast helps a film read as inhabited. Films from Arch2Video always include a small cast of original, fictional people in at least one scene, and people visible in your uploaded photos are not animated. Whatever tool you use, review the cast on two counts.

Plausibility and scale. Figures should walk at a human pace, keep the right number of limbs and fingers in every frame, and interact with the architecture correctly: feet on the floor, hands on the handrail, no bodies passing through glass or furniture. Pause as a figure crosses a threshold or passes a column; that is where errors cluster. A figure next to a door also sets the door's apparent height; a seven-foot person makes your door look small.

Identity. Confirm that no one in the film is a real, identifiable person who hasn't agreed to appear. If your source photos include a client or your own staff, use photos without people or confirm that the tool does not animate the people in the source image. Also check what the cast is doing: a crowd in a quiet chapel tells the wrong story about the program.

Lettering is the most reliably broken element in AI video: building numbers, wayfinding, a client's logo on a wall and book spines tend to render as plausible shapes rather than readable characters, and they shift as the camera moves. Review every piece of visible text at full size. If it is a smear, choose a source image that crops the signage out, mask it in the edit, or accept it only if it is small, out of focus and not a client's brand. The same applies to overlays: if the film carries your studio logo, check its proportions and edges, and a film that comes in logo and no-logo versions gives you a fallback if the overlay is wrong.

Scene order versus the story you want

A six-scene film has an implied narrative: approach, enter, move through, look closely, look back. Some tools let you set that order; others, including Arch2Video, use an AI director that plans the scene order itself, so the sequence may differ from your upload order. Either way, review it against what you want the film to say.

Does the film open on the image you would put on the cover? Does it move from exterior to interior the way a visitor would? Does the last scene leave the viewer with the project's strongest view? If not and you cannot reorder, a different set of six images may give the director a better finale. For unbuilt work, a film that moves from a hand sketch to a photoreal interior implies the interior exists, so make sure the caption says otherwise.

Audio, platform specs and disclosure wording

Music is easy to forget in a visual review and expensive to get wrong. Confirm that you have the right to use the track on every platform where you're posting: a track licensed for a website hero may not be licensed for Instagram, and platform-provided music is often restricted to personal accounts. Music for architecture videos covers the licensing categories. If your tool supplies an original soundtrack, as Arch2Video does with Still Air, Quiet Piano, Concrete or no music, keep a copy of the terms. Then confirm the music fits the project and that there is no unwanted audio; a bird call indoors is a reason to mute and replace.

Before export, match the film to where it will live. Vertical social platforms expect 9:16, and most AI film tools built for social output deliver 1080 × 1920 MP4, which is what Arch2Video produces; a website hero or a conference screen may need 16:9. Check that the file plays cleanly in the platform's own uploader, and watch the first two seconds: on autoplay feeds, that is the whole film for most viewers.

Decide how you will label the film before you write the caption. Say that motion was AI-generated from your images, and say whether the project is built: "AI-generated motion from our project photography; completed 2025" or "Concept study: AI-animated from design renderings, not yet built." AI rendering disclosure rules for architects has caption wording to copy and explains when platforms expect a label. Many platforms also attach their own AI labels from metadata in the file, so your caption may not be the only label the viewer sees; content credentials for AI video explains how that works and why it is better to write your own line than to let the platform's label speak for you.

The printable checklist

Print this and tick it off per film. Each row is pass or fail; one fail in the geometry, people or identity rows means the scene goes back.

What to do when a scene fails

The instinct is to rewrite the prompt. Most of the time the better move is to re-pick the image. An image-to-video model builds everything from the source frame; if the frame is soft, heavily compressed, cropped through a stair landing or lit differently from the other five images, no prompt will reliably fix the result. In a tool with no prompt box at all, the image is your only lever, and in practice it is the one that matters most.

1. Identify which scene failed and which row of the checklist it failed on.
2. Open the source image for that scene and look for the cause: low resolution, a crop through a key element, a composition that invites a pan, or a lighting mismatch with the rest of the set.
3. Replace it with a different photograph or rendering of the same view, or a tighter view that removes the problem element.
4. Keep the other five images unchanged so you can compare.
5. Regenerate and run the checklist on the whole film, since scene order and pacing can shift.

Generation typically takes an hour or more per film, so budget for one or two rounds rather than planning to post the first output the same afternoon. The film is an interpretation of your images; you are the one who knows what the building actually is.

Frequently asked questions

How long does a proper review of a 30-second AI architecture film take?

Plan on 20 to 30 minutes for a six-scene film: one viewing at speed, three paused frames per scene against the checklist, then five minutes on audio, specs and the caption. It gets faster once you know where your tool tends to fail.

What is the single most common failure?

Geometry on lateral camera moves: mullions and railings that bend during a pan or truck. The fix is on the input side, choosing images whose composition suits a static shot or push-in.

Can I fix a bad scene in a video editor instead of regenerating?

Sometimes. Cropping out a warped sign, trimming the last second where a line starts to bend, or muting unwanted ambient sound are reasonable edits. Bent verticals, texture drift and figure errors can't be edited out; those scenes need a new source image.

Should the checklist be different for unbuilt projects?

Add two checks: that the caption and any on-screen text present the project as a concept, and that the film does not imply a level of resolution the design hasn't reached. Unbuilt work tends to fail the materials and geometry rows more often because renderings and sketches give the model looser information than photographs.