0006 — Avatar render economics: one GPU clip per deck¶

Status: Accepted (backfilled 2026-09)
Drivers: wan-s2v/sadtalker take minutes per slide on consumer GPUs (an RTX 3090 renders a short clip in ~10 min, cold model load dominating); decks have 10–30 slides

Decision¶

settings.avatar treats GPU renders as scarce:

reuse_clip (default false): render ONE head clip from slide 1's voice and loop it under every later slide — one GPU render per deck, approximate lip-sync. Intended for cartoon mascots, where "talking loop" beats "correct phonemes"; do not combine with full_length: true (clip then matches only slide 1's narration length).
slides ("all" | N | [1, 5]): put heads only on chosen slides.
reuse_clip interplay: heads wanted but no narration audio → counted as a headless slide, never a silent stall.
ComfyUI hiccups between slides are survived and retried (commit 51b10a9); transient GPU OOM retries with backoff (oom_retries, oom_backoff), with a bounded poll timeout per clip.
No-GPU alternatives ship for the common cases: static (still image), puppet (loudness mouth-flap), precomputed (ready-made clips), and hosted D-ID as the BYO-key escape hatch.

A 20-slide mascot video costs the same GPU time as a 2-slide one when reuse_clip is on — the setting that makes mascots practical.
Sync is visibly approximate on reused clips; humans/photo-real avatars should pay for per-slide renders instead.