Setup: MacBook Pro M5 Max 64 GB · macOS 26.5.2 · vpipe v0.1.23 (also reproduced on a source build of b8c423b) · MiniMax-H3 FL2VA upstream BF16 DiT + BF16 text encoder (local diffusers layout) · larryvrh Turbo v4-600 runtime LoRA · 124 frames · steps=5 (4 NFE) · seed fixed.
Symptom: the final 4 pixel frames (121–124) carry a band of random colored macroblock noise across the bottom ~25 % of the frame. The rest of the clip is perfect.
Reproduced identically at BOTH orientations of the 768p canvas:
- 1344×768: corruption in frames 121–124, band from y≈575 down (≈192 of 768 rows)
- 768×1344: corruption in frames 121–124, band from y≈1020 down (≈320 of 1344 rows)
Same relative area (~bottom quarter), same frame range, deterministic across two runs → indexing/dispatch-size issue rather than a watchdog kill.
Controls:
- 544×960 × 124 frames, same seed/pipeline: clean through the last frame (checked frame-by-frame)
- 64×64 × 22 frames smoke: clean
- Mid-video temporal-chunk boundaries (frames 14–19, 31–36): clean — the defect is NOT periodic per chunk, it hits only the tail of the final chunk
- Detection method: per-frame JPEG size at fixed q (noise doesn't compress): frames 115–120 ≈ 66–76 KB, frames 121–124 jump to 88–131 KB
Numbers that may help locate it: 768p latent grid is 48×84 (4032 spatial tokens/latent frame) vs 544×960's 34×60 (2040). 4 corrupt pixel frames = exactly one latent frame at the 4× temporal compression; bottom quarter ≈ the tail of that frame's rows. Looks like the tail of the final chunk's output buffer is never written (uninitialized memory), with the threshold somewhere between 2040 and 4032 tokens/frame.
Happy to run any instrumented build or dump intermediate buffers — this machine reproduces it in ~12 min per run. (Same box as the residency-overcommit report; the DiT side works beautifully — 1344×768 · 4 NFE · BF16 in 11 min 37 s.)
Setup: MacBook Pro M5 Max 64 GB · macOS 26.5.2 · vpipe v0.1.23 (also reproduced on a source build of b8c423b) · MiniMax-H3 FL2VA upstream BF16 DiT + BF16 text encoder (local diffusers layout) · larryvrh Turbo v4-600 runtime LoRA · 124 frames · steps=5 (4 NFE) · seed fixed.
Symptom: the final 4 pixel frames (121–124) carry a band of random colored macroblock noise across the bottom ~25 % of the frame. The rest of the clip is perfect.
Reproduced identically at BOTH orientations of the 768p canvas:
Same relative area (~bottom quarter), same frame range, deterministic across two runs → indexing/dispatch-size issue rather than a watchdog kill.
Controls:
Numbers that may help locate it: 768p latent grid is 48×84 (4032 spatial tokens/latent frame) vs 544×960's 34×60 (2040). 4 corrupt pixel frames = exactly one latent frame at the 4× temporal compression; bottom quarter ≈ the tail of that frame's rows. Looks like the tail of the final chunk's output buffer is never written (uninitialized memory), with the threshold somewhere between 2040 and 4032 tokens/frame.
Happy to run any instrumented build or dump intermediate buffers — this machine reproduces it in ~12 min per run. (Same box as the residency-overcommit report; the DiT side works beautifully — 1344×768 · 4 NFE · BF16 in 11 min 37 s.)