Why does AI video warp buildings? Fixes for bending lines, melting furniture, flicker and invented rooms
Why AI video warps buildings: what causes bending verticals, melting furniture, flicker and invented rooms, the fix for each, and when to drop a scene.
In this article
You uploaded a clean rendering, and the clip that came back has a facade that leans as the camera moves, a sofa that softens into the rug and a window that quietly becomes two. The rendering isn't wrong, but it's usually where the fix starts. An image-to-video model predicts how pixels should move; it never sees your model, so straight lines and repeated elements are the first things to slip. This guide explains why AI video warps buildings, matches each common artifact to its likely cause, gives the fix that usually works, and tells you when to stop retrying and drop the scene.
Key takeaways
- AI video warps buildings because an image-to-video model has no geometry behind the picture: it guesses each new frame from pixels, and long straight edges, repeated windows and fine patterns show its errors first.
- Most artifacts trace back to three levers, in order of impact: the source image (resolution, perspective, clutter), the amount of camera motion, and the prompt.
- Correct converging verticals and wide-angle distortion in the still before you animate it, because the model extends whatever lean the image already has.
- Camera moves that stay inside the frame, such as a slow push-in, hold geometry far better than pans, orbits and pull-backs that force the model to invent what lies past the edge.
- Drop a scene rather than retrying when a structural element changes (window count, stair, roofline) or the clip shows a room that doesn't exist in the design.
Why AI video warps buildings
AI video warps buildings because an image-to-video model doesn't know your building is rigid. It has no 3D model, only the pixels of one image, and it predicts every new frame by guessing how those pixels should move, so its errors show up first in what architecture has most of: long straight edges, repeated openings and fine regular patterns. The fixes, in order of impact, are a better source image, less camera motion and a simpler prompt.
A rendering engine moves a camera through geometry, so a column stays a column from every angle. A video model generates each frame to look plausible next to its neighbors, not to match a plan. Plausible is a low bar for a tree and a high bar for a curtain wall: a slight wobble in foliage is invisible, while the same wobble in a mullion reads as a mistake.
It gets worse the further the camera travels from the view in your image, because the model has to invent more, and the longer the clip runs, because errors accumulate.
Match the artifact to its cause
Name what you see first; fixing the wrong lever wastes a generation.
| Artifact | What it looks like | Most likely cause | First fix | Drop the scene if |
|---|---|---|---|---|
| Bending verticals | Corners, columns or mullions lean or bow as the camera moves | Lateral move or orbit; converging verticals in the source | Correct verticals in the still; slow push-in or static camera | The lean is on the main facade after two retries |
| Melting furniture | Chair legs, cushions or fixtures soften, merge or change shape | Small, low-contrast objects in a soft or compressed image | Use the full-size original; crop tighter so key pieces are larger | A piece the client chose turns into something else |
| Flicker and shimmer | Brick, louvers, mesh, tile or grass crawl from frame to frame | Fine repeating patterns at the limit of the image's resolution | Higher-resolution source, or a view where the pattern is larger | The shimmer covers the main surface of the shot |
| Morphing windows | Window count, mullion spacing or door positions change mid-shot | Repeated elements the model can't count, passing the camera | Shorter move; a view with fewer repeated openings in motion | Openings change on a built or approved building |
| Invented rooms | A doorway opens onto a corridor or room that isn't in the design | Camera move past what the image shows | Choose a view whose move stays inside the frame | Always for built work; for concepts, if it contradicts the plan |
| Uncanny people | Faces distort, hands merge, walking looks weightless | Figures that are small or cropped in the source; people crossing key elements | Remove figures from the source images and let the tool add its own | A face is large in frame and wrong |
The people row has its own trade-offs, covered in choosing people for architecture videos.
Bending verticals and leaning facades
Bent lines are the artifact architects notice first, and they usually have two causes working together.
The lean is already in your image
A rendering from a low camera with a wide field of view has three-point perspective: the verticals converge toward the sky. In a still, the eye accepts it. In motion, the model re-estimates that convergence on every frame, and small disagreements make the building appear to breathe or tilt. Barrel distortion from a very wide lens does the same to horizontals near the frame edge. Fix it before you animate:
- Use your renderer's two-point perspective or vertical correction option, or correct the verticals in a photo editor.
- Narrow the field of view and move the camera back; a longer lens keeps edges straighter.
The camera move asks for depth it can't see
A pan, truck or orbit shows parallax: near objects slide past far ones. To fake it from one image, the model has to guess how far away every surface is, and every guess that's slightly off bends a line. A slow push-in mostly enlarges what's already in the frame, so there is much less to guess, and a static camera with moving clouds, foliage or water is safer still. Camera moves ranked by how well geometry survives goes through each move. If a lateral move is essential, shorten it.
Melting furniture and drifting materials
Furniture melts when the model can't read what it's looking at. A small chair in a dark corner, against a rug of similar color, is redrawn from a guess on every frame, and the guesses don't agree. Materials drift for the same reason: a soft, compressed image of oak flooring gives the model too little grain to hold. The fixes are mostly about the file:
- Use the original, full-size export. A copy sent through a messaging app or pulled from a website is compressed and often downscaled. For Arch2Video, images need a long side of at least 512 px, and 1,500 px or more gives the sharpest scenes; the principle holds for any tool (export settings for renderings headed to video).
- Make the important pieces bigger in frame, and simplify the styling. A tight crop of the living area holds the sofa better than a wide shot of the floor, and shelves of small accessories rarely survive.
- Keep motion away from key pieces. Melting often follows something passing in front of an object, because the model has to redraw it from memory.
Flicker and shimmering textures
Shimmer is a resolution problem that motion exposes. Brick coursing, perforated metal, battens, louvers, tile grids and distant rows of windows are fine repeating patterns. When a pattern is close to the smallest detail the image can resolve, the model can't decide where each line falls, so it crawls from frame to frame.
Before blaming the model, check the file you downloaded, not the post on your phone. Social platforms re-encode uploads, and compression adds its own shimmer to fine patterns. If the original shimmers too:
- Use a higher-resolution source, so the pattern is comfortably larger than the image's finest detail.
- Choose a closer view: near brick reads as texture, distant brick reads as noise.
- Slow the camera down; fast motion across a fine pattern is the worst case.
- Go easy on sharpening, which adds halos the model then has to hold steady.
Glowing windows or fixtures pulsing at night behave differently; day-to-night sequences covers them.
Morphing windows and invented rooms
These two matter most, because they change the design rather than degrade the image.
Windows the model can't count
A facade with a dozen identical panes is, to the model, a pattern rather than twelve objects. As the camera moves, the pattern can gain or lose a pane, mullion spacing can shift and doors can slide along a wall. Keep shots of long repeated facades static or nearly so, prefer views where the openings are large in the frame, and keep moves short. If the window rhythm is the point of the design, give it a gentle push-in and nothing more.
Rooms that aren't in your plan
When the camera pushes toward a doorway, turns a corner or pulls back past the frame edge, the model has to show something it was never given, so it produces something plausible: a hallway, a second bedroom, a terrace. That is how these models work, not a malfunction. Choose images where the destination of the move is already visible, and avoid views composed around an open door or an unseen stair. A floor plan won't steer the model either (what AI does with a floor plan).
Prompt fixes that help, and the ones that don't
If your tool has a prompt box, keep it short and about motion. The image already shows the model what's there; what helps is saying what moves and how much. Use one camera move per clip, name it plainly ("slow dolly in toward the entrance") and ask for slow.
If there is a negative prompt field, list what you don't want rather than giving an instruction. Google's Veo prompt guide marks "using instructive language or words such as 'no' or 'don't'" as not recommended and says to "describe what you don't want to see" (Google Cloud documentation, updated October 28, 2026). So "bent lines, distorted windows" goes in the negative field, not "don't bend the lines." Other tools handle negatives differently.
What rarely helps: "architecturally accurate" or "keep the geometry perfect," which the model can't check, and prompts that stack moves, weather and people.
Some tools have no prompt at all. In Arch2Video the AI director plans the scenes from your six images, with no prompt box or editing timeline, so the image set is your only lever, and it's the one that matters most anyway.
When to drop a scene instead of fixing it
Retrying costs credits or time, and some scenes don't get better. A practical rule: if two retries with a better source and a calmer move still distort the same element, the image is the problem. Replace it. Drop the scene without further retries when:
- A structural element changes on a built or approved project: window count, roofline, stair, column spacing.
- The clip shows a room, opening or view that isn't in the design.
- A material or fixture the client selected turns into something else.
- A face is large in frame and visibly wrong.
An artifact can usually stay when it sits in foliage, sky or deep background, lasts a few frames and doesn't touch the subject. Watch it full screen on a phone, and run the whole film through a review checklist before you publish.
A good replacement is often a tighter view of the same space: a detail of the stair rather than the whole stair hall. In Arch2Video, each film uses each of your six images once, so fixing a scene means swapping that image and making a new film, which is a new video at USD 15. Whatever the tool, present unbuilt work as concept work; the motion is an interpretation.
Frequently asked questions
Why do straight lines bend in my AI video when the rendering looks perfect?
A perfect still isn't always easy to animate. Converging verticals, a wide field of view or a long regular facade look fine as a still and become hard to hold once the camera moves, because the model re-estimates perspective on every frame. Correct the verticals, narrow the lens and ask for a slow push-in or a static shot before changing anything else.
Will upscaling my rendering stop the warping?
Only if the warping comes from a soft or small source, and re-exporting at full size beats upscaling a small copy, because AI upscalers can invent detail that the video model then animates. Upscaling does nothing for bending caused by camera motion; that needs a calmer move or a different view.
Why does the AI add a room or a door that isn't in my design?
Because the camera moved toward something the image doesn't show, and the model filled the gap with something plausible. It can't see your plans, so it can't know the corridor is a closet. Choose images where everything the move reveals is already in the frame.
Is a small artifact ever acceptable to publish?
For a social post, a brief shimmer in a tree or a soft patch of sky usually is, if it doesn't touch the building. For a client presentation, a public hearing or a listing of a real property, anything that changes the design, adds a space or misstates a material should come out, and unbuilt work should be labeled as concept work.