Addaly is in open beta. Things will change, and AI answers can be wrong — check anything that matters.

Making Things With AI

Images, video, voice and music — how they work, where they break, who owns them.

Lesson 43 of 848 min

Why clips get worse as they get longer

The compounding problem

The obvious way to make a long video from a short-window model is to generate a chunk, take its last frames, and use them to condition the next chunk. Repeat.

It works, and it degrades in a predictable way. Each generation introduces small errors — a slight colour shift, softening, a face marginally off. Those errors are then treated as the truth for the next chunk, which adds its own on top. After five or six extensions the result has drifted somewhere the first chunk would not recognise.

The characteristic progression, which you can watch happen:

  1. Contrast and saturation creep upward, because each pass pulls slightly toward the model's mode.
  2. Detail softens, because each pass through the autoencoder loses a little.
  3. Faces regularise toward the average face.
  4. The scene reorganises, because nothing outside the current window constrains it.

This is the same mechanism as repeated image-to-image from the control module, running automatically, several times, inside one apparently simple feature.

What six chained extensions do to a clip05010016Extension numberMeasured against the first chunk—— Saturation, % above the first chunk– – Fine detail retained, %Each chunk is conditioned on the last chunk's errors, so the errors become the ground truth.Correcting the anchor frame by hand at extension two means regenerating one chunk; noticing at sixmeans regenerating four. Put the first and last frames side by side after every extension.
What six chained extensions do to a clip05010016Across: Extension numberUp: Measured against the first chunk—— Saturation, % above the first chunk– – Fine detail retained, %Each chunk is conditioned on the last chunk'serrors, so the errors become the ground truth.Correcting the anchor frame by hand at extension twomeans regenerating one chunk; noticing at six meansregenerating four. Put the first and last framesside by side after every extension.

Why long-video demonstrations are usually cut

When you see a minutes-long generated video, look at the cutting. Almost always it is many short generations edited together, with cuts placed exactly where continuity would have broken. That is legitimate filmmaking and it is worth naming, because it is frequently presented as a single capability.

The exception is the current class of causal streaming models, which generate frame by frame and can run indefinitely. They face the same drift and manage it with explicit mechanisms — periodically re-anchoring on a reference frame, or keeping a compressed memory of earlier content. These reduce drift; nothing yet removes it.

Working with it rather than against it

Re-anchor deliberately. Rather than chaining chunk to chunk, condition every chunk on the original reference image as well as the previous frames. Drift is then measured against a fixed point rather than against the last error.

Regenerate the anchor by hand. If chunk three's last frame has drifted, fix that frame as an image — inpaint the face, correct the colour — and use the corrected version to condition chunk four. Ten minutes of image work resets the accumulation.

Grade at the end, not per chunk. Colour-correcting each chunk to look good individually makes the drift harder to see and no smaller. Assemble everything, then grade the whole sequence to one reference. Free tools do this well; Resolve's free edition is the obvious one.

Cut on motion. A cut during movement hides a mismatch that would be obvious across a held frame. This is standard editing practice and it is doubly useful here.

The arithmetic worth doing before you start

Before committing to a piece, work out how many independent shots it needs and how long each is. A two-minute explainer is not two minutes of generation; it is perhaps thirty shots of three to six seconds, most of which will need several attempts.

That number tells you whether the project is feasible on your budget and your time, and it is far better to know at the start. It also usually reveals that half the shots do not need generation at all — a title card, a screen recording, a still with a slow move on it, a piece of stock footage. Mixing sources is not a compromise. It is how the medium is actually produced.

A drift check that takes ten seconds

Put the first frame and the last frame of your extended clip side by side and look at them as a pair, not as the ends of a sequence. Watching the video hides drift, because your eye follows the change; seeing the two frames together makes it obvious.

Do this at every extension rather than at the end. Drift is far cheaper to correct at chunk two than at chunk six, because correcting it at chunk two means regenerating one chunk and correcting it at chunk six means regenerating four.

The same trick works for a sequence of separate shots that are supposed to match. Lay the first frame of every shot in a row in any image editor. Mismatched grade, drifting colour and inconsistent contrast are visible instantly in that strip and invisible when the shots play in order.

The honest limitation: there is no setting that makes a long generated take hold together. Length is bought with cuts, anchors and manual correction. Anyone promising a single continuous generated scene of any duration is either using a very forgiving subject or has not looked closely at the result.

The one thing to keep

Extending a clip by conditioning each chunk on the previous one accumulates error, because each chunk's small imperfections become the ground truth for the next.

Before you move on

A five-chunk extended clip ends up brighter, glossier and softer than it began. What is the mechanism?

Pick the one you would defend. Nobody sees your answer.

No ads. No data sale. No public scores on people. Ever.

© 2026 Addaly