# Why AI video looks warped — bending walls and sliding limbs > Bending architecture, sliding limbs and melting objects share one cause — the engine has no 3D model of your scene. Here is what actually reduces it. Published 2026-09-05 · eroq.ai — canonical: https://eroq.ai/blog/why-ai-video-looks-warped The wall behind your subject bows like a funhouse mirror. An arm passes behind a table and comes back as two arms. A cup slides across a counter it was never touching. Nothing errored, nothing was refused, the clip rendered exactly as asked — and it is unusable. Warping is not a bug you can file. It is the shape of the technology, and once you know why it happens you can steer around most of it. (If your render did not arrive at all, that is a different problem with its own drill, covered in [why AI video renders fail](/blog/why-ai-video-renders-fail).) ## The engine has no idea what a table is A video model does not build your scene, light it and photograph it. It predicts frames. There is no 3D geometry underneath, no skeleton on the figure, no object that persists across time and keeps its identity. What exists is a very good guess about what the next frame should look like, given the previous one and your prompt. That guess is excellent while the picture stays in territory the model has seen a million times — a face front-on, a slow push-in, a body walking in a straight line. It falls apart where the model has to invent geometry it never observed: a hand crossing in front of another hand, a chair leg disappearing behind a leg and coming back, a reflection that has to agree with the thing reflecting it. The engine does not know those constraints exist. It renders the most plausible continuation, and plausible is not the same as consistent. Every fix below is therefore the same fix wearing different clothes. Give the model less to invent. ## One action per clip The most reliable cause of melted footage is a prompt that describes a sequence. "She picks up the glass, turns toward the window, and the camera pulls back as the lights come on" is four events in five seconds. The engine attempts all four, blends them, and hands you a person who is simultaneously lifting and turning while the room dissolves behind her. Write one action. If the shot needs the glass *and* the window, that is two clips and a cut. Film mode in [the video studio](/studio/video) exists for exactly this — a storyboard where each scene renders on its own and plays back assembled. The same rule governs the camera. Two camera moves in one prompt fight, and only one survives. The composer reads written directions out of your text and sets the matching parameter, one per kind — and which of two competing moves wins is decided by the lexicon, not by the order you typed them. Write "slow push-in" or "handheld", never both, and let the [director's rack](/glossary/director-rack) carry the rest of the look. ## Shorter clips bend less Instability compounds over time. A shot that is clean at three seconds often starts drifting at eight, because every frame inherits the errors of the one before it. When a shot warps, halve it before you rewrite it. This is also the cheap way to find out whether the prompt or the engine is at fault. A three-second draft on [Seedance 1.0 Lite](/models/seedance-1-lite) costs 36 credits — clip pricing runs off a five-second anchor plus a marginal rate per second beyond it, so short really is cheaper rather than merely proportional. If the three-second version holds together, your prompt is sound and length was the problem. If it warps too, you are describing something this engine cannot build. ## Pick the engine that can do the physics Engines are not interchangeable, and a lot of "this looks warped" is a shot asked of the wrong one. - **Bodies in motion and faces that have to stay faces** — [Kling 2.5 Turbo](/models/kling-2-5-turbo), which is also the only video engine here with a native negative prompt, so "no extra limbs, no distorted hands" has somewhere real to go. - **Weight, falling, splashing, anything that has to obey gravity** — Hailuo 02, whose entire personality is cinematic physics and dramatic light. - **Long takes that must stay coherent** — [Seedance 2.5](/models/seedance-2-5) runs 4 to 30 seconds and holds its geometry across that span better than engines built for five. - **Mature material** — Motion One, always. The premium engines are moderated by their providers and will refuse rather than render something compromised. Trying an engine your plan does not include is free information rather than a charge — the gate answers before anything is billed. ## Anchor the subject with a frame The single most effective change is to stop asking the engine to invent the subject at all. Render the still first, approve it, then animate it. In [image-to-video](/glossary/image-to-video) the first frame is fixed, so the model is no longer guessing what your character looks like — only what they do next. That removes an entire category of drift. Three ways to hand it that anchor: a reference photo on the composer, a [Character](/studio/characters) whose photos carry the same face across shots, or **Animate** on any image in your Library. All three Seedance engines go further and accept an end frame as well as a start frame, interpolating the journey between them. A loop, a reveal, or a move that has to land on one specific image stops being a gamble. ## A worked example Here is a shot written badly, paraphrased: a woman walks through a market while vendors move around her, the camera orbits, she picks up a fruit, golden hour, anamorphic, handheld, 35mm, a crowd. Four actions, three camera instructions, an unspecified number of extras. The fix is one paragraph with one of everything: > A woman in a linen shirt stops at a market stall and turns a peach over in her hand, studying it; slow push-in from a medium shot to a close-up on her hands, late golden-hour light raking across the stall, dust in the air, a calm and unhurried mood. Forty-nine words. One subject, one action, one camera move, one light source, one mood. The vendors and the orbit become a second clip. ## What will not fix it Adding adjectives. "Photorealistic, highly detailed, consistent anatomy, no artifacts" spends your prompt on instructions the engine has no mechanism to follow, and dilutes the words that do steer it. And the honest limit: there is no repair step here. eroq has no inpainting, no upscaler, no face swap and no editing timeline, so you cannot paint out a bent doorframe. The fix for a warped clip is always another render — which is exactly why the cheap three-second diagnosis matters more here than it would in a tool that let you patch the output. Download the keeper and take it to Premiere or After Effects when it needs real post. [Open the video studio](/studio/video) and try the three-second version first. ## FAQ ### Why does architecture bend when the camera moves? A moving camera asks the model to reveal parts of the scene it has never rendered, and it has no floor plan to reveal them from. Straight lines are where that invention is most visible, which is why walls, doorframes and window mullions bow first. Slower and shorter moves give it less new geometry to invent. ### Does rendering at a higher resolution reduce warping? No. Resolution changes how many pixels the artifact is drawn with, not whether the artifact happens, so a 1080p render of a bad shot is a sharper bad shot. Clip length, the number of subjects and the engine's fit for the motion are what actually move the needle. ### Can I fix a warped clip without re-rendering it? Not inside eroq. There is no inpainting, no upscaler and no editing timeline here, so the only fix is another render — usually shorter, with one action, or anchored by a start frame. Download the clip and take it into Premiere or After Effects if it needs genuine post work. ### Does a negative prompt help with warping? Only on Kling, which is the one video engine with a native negative prompt. Everywhere else the field is accepted and then quietly ignored by the engine, so it is not doing what you think it is doing. On the other engines, spend that effort on a shorter clip and a start frame instead.