A real, technically-verified H3 render — sitting past its own contest deadline
Another gap the 2026-09-05 full-coverage check missed:
it audited blades68-lora name-by-name and T2VA for all but one branch
(stock-turbo-lora-baseline, entered as Blocked). T2VA's feature/h3-sync-sound-challenge was
the second miss — real work, no write-up, and a time-sensitive detail nobody had flagged: its
target deadline is now five days gone.
What it was for
Gavin flagged the official Comfy H3 Sync Sound Community Challenge
(blog.comfy.org, checked
against the primary post directly), deadline 2026-09-01. Hard rule from the post itself: audio
must be generated natively by H3 in the same pass, no post-hoc stitching. T2VA already had a
production graph (build_t2va_prompt.py) exercising H3's audio-capable
MiniMaxH3ReferenceToVideo path — confirmed for real (non-silent) audio via ffprobe on an
existing production render before this branch even started. The entry itself: a short Raven
tactical vignette (turn, weapon-ready foley beat, one line of dialogue) built via a new
build_sync_sound_prompt.py, parameterizing clip length instead of production's fixed 107 frames.
What actually ran
v1 dispatched through Concourse's isolated t2va-sync-sound-challenge pipeline (build 1,
succeeded 2026-08-27, 22m26s). Technical verification followed 2026-08-28, pulled from the actual
comfyui-local output file rather than a claimed status line:
- h264, 480x864, 158 frames / 6.583s — exact match to the spec's own length target
- Real stereo AAC audio, 32kHz, mean -20.1dB / max -0.5dB: non-silent, not clipped
- Confirmed live in Immich (album "Esoteria T2VA Character Sheets", asset
28602f82-...)
Every automatable check passes. What's explicitly not automated, per the branch's own
EXPERIMENT.md: whether the footstep/turn sound actually lands on the pivot motion, whether the
weapon-lock sound lands on the weapon reaching ready position, and whether the dialogue line's
mouth movement reads as lip-synced. That's Gavin's call by design, not a gap in the harness.
The part nobody flagged
That qualitative review never happened — the branch's last substantive commit is the 2026-08-28
verification log. Checked directly rather than assumed: no commit anywhere in this branch's history
or T2VA's broader log mentions "submit" or "challenge" past that date, and the contest's own
2026-09-01 deadline is five days behind today. Whether Comfy's own submission process needed
something beyond "render exists in Immich" (a form, a forum post, an upload) isn't something this
repo's history can answer — but if it did, that window has closed with a technically-sound entry
sitting unsubmitted and unreviewed.
Verdict: Pending, and now moot on the original timeline. The render itself remains a real, verified data point on H3 native audio-sync feasibility regardless of contest outcome — it's the "did we actually enter" question that lapsed silently.