Repository navigation
First release. Unofficial ComfyUI integration of NVlabs LongLive-Plug adapters for Wan2.1-T2V-14B.
What's included
LongLive-Plug Apply Adapter Pair: applies the official Few-Step (1.0) + CFG (0.5) LoRAs and checks them: backbone fingerprint, PEFT/native key mapping, rank/alpha, shapes, finiteness, unused keys, pinned release SHA-256. It patches a model clone and returns a JSON coverage report (400/400 targets per adapter).LongLive-Plug Sampling Recipe: GUIDER/SAMPLER/SIGMAS for the LongLive-Plug 4-step FlowUniPC recipe, bit-identical on CPU to the upstream scheduler. Also included: the two other 4-step step-list recipes published with the adapter, and the official Wan2.1 50-step baseline.- GUI and API workflows (baseline, 4-step), comparison runner, merge verification script, model-free CI (Ubuntu and Windows).
Tested
Windows 11, RTX 5090 32 GB, ComfyUI v0.38.0, PyTorch 2.14.1+cu130, Python 3.12, bf16 Wan2.1-T2V-14B (Comfy-Org repackage), ComfyUI started with --bf16-unet --bf16-text-enc --fp32-vae, 832×480×81 frames. Other OSes, GPUs, quantized bases and other Wan variants are untested or refused; see the README status table.
Measured on that machine (same prompt, seed 20261002)
| Run | Model passes | Wall time |
|---|---|---|
| LongLive-Plug 4-step (first run / second run) | 4 | 59.0 s / 44.5 s |
| Wan2.1 base 50-step CFG 5 (first run after server start) | 100 | 943.5 s |
| Base, naive 4-step CFG 5 (comparison) | 8 | 84.9 s |
Method, memory notes and caveats: docs/BENCHMARKS.md.
Assets
longlive_plug_4step_seed20261002.mp4: LongLive-Plug, 4 steps (832×480, 81 frames, 16 fps)baseline_50step_seed20261002.mp4: Wan2.1 base, official 50-step recipenaive_4step_seed20261002.mp4: base, 4 steps without adapters (comparison)run_manifest.json: raw run records;SHA256SUMS.txt: checksums
Demo prompt written for this release. Videos generated by the maintainer with Wan2.1-T2V-14B (Apache-2.0) and the LongLive-Plug adapters (Apache-2.0). Not affiliated with NVIDIA, the Wan team or Comfy Org.