pipeline.py: _encode_latents must NOT divide by vae scaling_factor. The
VideoMaMa UNet is trained on raw VAE latents (matching okdalto's
VideoInferencePipeline, which multiplies by scaling_factor then divides it
back out -> net x1). Dividing over-scaled the conditioning ~5.5x and produced
a systematically blurry matte on every frame. Verified 5/5 vs the desktop /
vanilla reference on identical single-frame inputs.
nodes.py: harden the loader precision fallback so fp16 is the true default -
only an explicit "bf16" opts out; a stale "auto" widget now resolves to fp16
instead of bf16.
sync_to_dev.py: the rmtree+copytree mirror deleted matte/checkpoints/ (the
gitignored ~10GB SVD+VideoMaMa weights) on every matte sync, forcing a
re-download. Add PRESERVE_IN_DEST + selective _clean_dest so runtime-
downloaded model data survives a sync.
docs: add docs/maintainers/matte-pipeline.md (latent-scaling invariant,
black-output causes, fidelity levers, single-frame + vanilla-oracle debug
loop, checkpoint preservation).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
