The Artifact Atlas
ManageableFixable with known levers.

Temporal flicker

Frame-to-frame instability in brightness, texture or fine detail that is invisible in a still and obvious the moment the clip moves.

Also called

  • Boiling
  • Shimmer
  • Frame jitter
  • Crawling detail
Evidence

Watch the surface detail rather than the subject. Any single frame here would pass review; the sequence does not.

Causes
  1. 01Frames generated with insufficient temporal conditioning, so each one reconstructs high-frequency detail independently.
  2. 02Noise re-initialised per frame instead of held consistent across a sequence.
  3. 03Latent windows that do not overlap enough, leaving each window free to resolve texture its own way.
  4. 04Decode variance — the same latent can decode to visibly different fine detail run to run.
  5. 05Upscaling or detail-enhancement applied per frame after generation, which amplifies whatever inconsistency already existed.
What actually helps
  1. 01Hold noise consistent across the sequence rather than re-seeding per frame.
  2. 02Increase overlap between latent windows and blend across the overlap.
  3. 03Move texture and grain intent to a single post stage so it is applied once rather than re-synthesised per frame.
  4. 04Reduce reliance on high-frequency detail in the style itself — fine texture is the first thing temporal instability destroys.
  5. 05Deflicker as a last resort. It treats the symptom and softens real detail along with the artifact.

Tends to originate at

  • Per-frame sampling without temporal conditioning
  • Latent window boundaries
  • Decode stage

How to confirm it

Step through frame by frame. If each frame is individually plausible but the detail is in a different place each time, it is flicker rather than a bad generation — and the fix is upstream of anything aesthetic.

The useful diagnostic is to isolate a static region. Find part of the frame that should not be changing at all — a wall, a flat background, an unmoving surface — and watch only that. Motion hides flicker; stillness exposes it.

Why it is worth naming separately

Flicker is routinely misdiagnosed as “the model isn’t good enough”. It usually is not a capability problem. It is a conditioning or scheduling problem, and it often has a cheap fix that has nothing to do with switching models.

It also has an unusually asymmetric relationship with review: it is nearly invisible in the still frames most approval processes actually look at, so it survives review and then fails in front of a viewer.

Related