Datamosh
Datamosh is intentional codec failure: inter-frame prediction continues after keyframes are removed or broken, so motion vectors smear the previous picture into the next GOP and faces bleed into the following shot. The image is file wreckage, not a lens error.
Narrative function
The I-frame never arrives. P-frames keep describing motion that no longer has a valid picture.
A mouth stretches into a wall. A cut happens in time but not in pixels.
Digital ruin is the style. Use it when the file should break on purpose: a music video, a corrupted memory, a broadcast that fails.
The smear should follow motion vectors, not a random RGB offset. It is the wrong tool for a stutter of whole frames (that is stop-stutter), for chromatic fringing, or for an altered-state smear from a long shutter.
If the image trails because the shutter was open, that is photography. If it trails because the codec predicted the wrong future, that is datamosh.
How it works
Encode with inter-frame compression, drop or repeat I-frames, let P-frames or B-frames run. The artifact is temporal and blocky, often with residual macroblocks.
Prompt: inter-frame collapse, faces smearing into the next GOP, codec wreckage, not RGB split, not film smear.
When to use it
A cut from a face to a city in which the face's pixels are dragged across the skyline by leftover motion vectors. Do not step-print the same clip or add red-cyan edge fringe and call it datamosh.
Compared with similar shots
Stutter / Stop-Stutter repeats or steps whole frames as a temporal rhythm. Datamosh destroys the picture between frames.
Altered State is perceptual and photographic. Chromatic Aberration is optical color fringing.
Datamosh is codec wreckage, not a general noise overlay.
In film
Chairlift's Evident Utensil is a well-known datamosh music video: GOP collapse as the picture. Other music videos adopted the same inter-frame smear.
Features rarely hold it long because it destroys coverage.
Prompt it
Break prediction, not the lens. Still: Datamosh of [Subject], inter-frame collapse, faces smearing into the next GOP, digital ruin as style.
Video: Datamosh corruption around [Subject], frames bleeding into the next, codec wreckage as image, not a shutter smear.
What usually goes wrong
Wrong cousin: stutter, long-shutter smear, or RGB split. Tell: clean frame repeats, photochemical motion blur, or three colored copies of the subject.
Fix: blocky temporal bleed across a cut, leftover pixels from the previous picture, no new I-frame.
Questions
What is datamosh in film?
Deliberate video-codec collapse. Without a fresh keyframe, motion vectors smear the last valid picture into the next, producing inter-frame wreckage.
Datamosh vs stutter
Stutter repeats or steps complete frames. Datamosh lets predicted frames destroy the image. One is a rhythm of whole pictures. The other is a melt between them.
When should I use datamosh?
When the file should fail as the point: music video, corrupted feed, a digital breakdown. Do not use it as a substitute for optical smear or a simple jump cut.
How do I prompt datamosh for AI image or video?
Ask for inter-frame collapse, GOP smear, codec wreckage across a cut. Forbid shutter blur, RGB channel split, and freeze-frame stutter unless those are extra.
Melies can generate this Datamosh
Melies is the AI filmmaking workspace. This Datamosh opens with the optical prompt loaded, so the effect has a physical cause in the frame.