Add two opt-in inference speedups to the Model Loader, validated to leave the
output perceptually identical (deviation at the fp32 rounding floor):
- tf32 (default on): TF32 matmul on Ampere+ (~1.15x).
- compile (opt-in): torch.compile the UNet (~2.1x). Stacks with TF32 to
~2.5x (measured 4.3s -> 1.7s on a 12s clip).
torch.compile needs a static shape (the model's adaptive-avg-pool can't trace
dynamic shapes), so the sampler pads every chunk to chunk_seconds — clips of
any length reuse one compiled graph (no per-length recompiles; verified an 8s
clip after a 12s clip ran in 0.9s with no recompile).
Researched + profiled first: CFG-batching, channel/chunk batching, and
channels_last gave ~0 gain because the GPU is already compute-bound at batch 1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ComfyUI nodes for UniverSR (ICASSP 2026) — vocoder-free audio
super-resolution (8/12/16/24 kHz → 48 kHz) via flow matching.
- UniverSR Model Loader: presets auto-download to models/universr,
plus local dir / raw .pth (from_local) loading, with caching.
- UniverSR Super-Resolution: chunked overlap-add for long audio,
per-channel stereo, seed control with global-RNG isolation,
wet/dry blend, and an optional before/after spectrogram.
- Vendors the universr inference package under vendor/ (prefers an
installed copy); only extra dep beyond ComfyUI's stack is torchdiffeq.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ethanfel