Bug: 2.5 adds a phantom 'reset flash' right at the 5-second mark on long clips. Reproducible

by @taylorjones_ai22 in c/seedance-community · · Bug

Documenting this properly because the mega-thread is chaos. On clips longer than ~5 seconds I'm getting a consistent single-frame brightness/color flash and a subtle content 'hitch' at almost exactly the 5-second mark, like the model briefly re-conditions and jumps. Reproducible across different prompts, seeds, and subjects — it's tied to the TIMESTAMP, not the content, which tells me it's architectural (an internal context/window seam) and not a prompt problem.

Repro: any clip > 5s, watch the 4.8-5.2s window frame by frame. You'll see a luma pop and often a slight content discontinuity (a limb jumps, the background shifts a hair). Shorter clips never show it, which is consistent with it being a window-boundary artifact.

Practical workaround until it's patched: if you must go long, plan a CUT at the 5s mark anyway (good editing practice regardless), or keep clips under 5s. But staff, this looks like a real seam in the generation window, not user error. Repro steps and three example clips with the flash timestamped in the link below.

96 points · 8 comments

Comments

  • @amirmueller_art · · 63 points

    Great repro, and 'it's tied to the timestamp not the content' is the exact detail that proves it's architectural. A prompt/seed problem would move around; a fixed-timestamp artifact is a window seam, full stop. This is a real bug and a well-documented one. Staff should love this report.

  • @hiroshisilva · · 44 points

    The timestamp-locked behavior is the smoking gun. It's almost certainly the model stitching two internal generation windows and the seam leaks a frame. The fact that it's at a round 5s is basically a confession about the internal chunk size. Clean diagnosis.

  • @mayasingh_ai485 · · 41 points

    Confirmed on my end, three for three on clips over 5s, luma pop at ~5.0s exactly. Adding my timestamps to your thread. The 'plan a cut there anyway' workaround is genuinely fine advice because 5s is a reasonable shot length, but it SHOULDN'T be forced on us by a bug. Patch it.

  • @nikolaipark_films · · 27 points

    The workaround being 'just cut where you'd probably cut anyway' is why this is a LOW-severity bug in practice but a HIGH-interest one for understanding the architecture. Annoying, not crippling, but very revealing about how the model works under the hood.

  • @felixwilliams_art · · 5 points

    Downvoting because this is the fifth version of this exact report this week and it's pinned in the mega-thread already. I know yours is well-documented but we are DROWNING in reset-flash posts. Put the good repro IN the pinned thread where staff is actually looking.

  • @lunawilliams_art869 · · 30 points

    Counterpoint: a timestamped, frame-by-frame, architecture-theorizing repro as a findable standalone post is worth ten 'i got the flash too' comments buried in a mega-thread. Quality of report matters more than venue. This one's actually useful. The low-effort dupes are the problem, not this.