AI-generated video melts or morphs between frames because diffusion models generate each frame with limited cross-frame consistency. Every frame is a plausible image, but the model has no exact memory of object structure from one frame to the next, so geometry drifts — a hand grows a finger, a face slides, a sign smears. It reads as "melting" because large regions change while their crisp edges dissolve, rather than an object cleanly moving across the frame.
What's actually happening
A video diffusion model turns noise into a frame that matches your prompt. To make a clip, it produces a sequence of these frames with some mechanism to keep them consistent — but that consistency is approximate, not exact. The model knows "there should be a hand here, roughly like the last frame," not "this is the same hand with these exact five fingers in these exact positions."
So across a span of frames, structure drifts. Fingers merge or split. A face's proportions slide. Text on a sign turns to gibberish and back. The effect is strongest where the content is detailed and moving — hands, faces, small text, busy backgrounds — because that's where small per-frame errors accumulate into visible deformation.
Why it looks like melting, specifically
There's a clean way to describe the difference between normal motion and a morph, and it's the basis for detecting one:
- Real motion is edge-preserving. When an object moves, a lot of pixels change — but the object's crisp edges move with it and stay sharp. High change, high edge detail.
- Morphing is edge-dissolving. A large region changes AND the high-frequency detail — the edges — collapses into a blur instead of translating. High change, collapsing edges.
That combination — a lot changing while structure dissolves, sustained over a span rather than a single frame — is the signature of a melt. It's why a morph looks liquid: the picture is losing its edges, not moving them.
How to inspect a clip for it
UploadCheck's ai_warp gate inspects consecutive frames for exactly that signature — high change magnitude with collapsing edge detail, sustained over a span — and flags the timestamps where it appears. Point it at the finished export:
uploadcheck check render.mp4 --checks ai_warpA flag tells you the span to regenerate. Because morphing is usually localized to one shot, you re-prompt or re-roll just that shot instead of the whole clip, then re-check.
ai_warp inspects a finished file for the morph pattern above and reports where it finds it. It does not claim to detect that a clip was AI-generated, and it does not guarantee it catches every morph — subtle or edge-preserving deformation can pass. It is tuned to be conservative: on a 40-clip corpus of real, good AI-generated video it produced zero false positives, so it does not cry wolf on clean output. Treat it as a strong first-pass inspection, not a guarantee.
Morphing vs flicker vs frozen frames
Generated clips fail in a few distinct ways, and they don't share a fix, so UploadCheck checks them separately:
- Morphing (
ai_warp) — structure smears between frames. Fix: regenerate the shot. - Flicker (
flicker) — the whole frame's brightness pumps up and down. Fix: stabilize exposure or regenerate. - Frozen / looped frames (
loop_freeze) — a held or repeated frame where motion should continue, common at 8-second stitch seams. Fix: re-cut or regenerate the seam.
Catch it before your audience does
Morphing is the classic "looks fine at a glance, falls apart on the second watch" defect — which is exactly the kind that slips past a quick review and reaches your audience. Inspecting the export with ai_warp (on its own, or as part of the ai_render_gate pipeline gate) turns "hope nobody notices" into a timestamped work order before you publish.