TurboWan Model Loader (Quantized)

This is the node that makes the "100x faster Wan video" claim real - and it's also where most people hit their first error. TurboWanModelLoader loads the int8-quantized .pth checkpoints from THU-ML's TurboDiffusion project so the rest of the pack can sample Wan 2.2 in 4 steps instead of 20-30. In practice you'll run two of these, one for each of the model files the pack needs. More on that below.

What it's actually loading

Wan 2.2 isn't one model, it's a two-expert MoE: a high-noise expert that handles motion and composition, and a low-noise expert that refines detail. TurboDiffusion distills that pair into an rCM (rectified consistency model) that denoises in 1-4 steps - a trick in the same family as LCM and Lightning, applied to video instead of images. That's the speed, and it's also why the quality tradeoff feels familiar: distilled models are great for iterating, less great as the final word.

The files you want from the TurboDiffusion HuggingFace repo are:

• TurboWan2.2-I2V-A14B-high-720P-quant.pth (≈14.5 GB, int8)
• TurboWan2.2-I2V-A14B-low-720P-quant.pth (≈14.5 GB, int8)

Drop both into ComfyUI/models/diffusion_models/. The -quant files are block-wise int8; the full-precision .pth versions exist but weigh ~28.6 GB each and buy you little. (There's a verify_models.py script in the repo if you want to double-check your downloads.)

How it works

Under the hood the loader uses TurboDiffusion's own create_model() path (vendored inside the pack - no separate TurboDiffusion pip install), builds the architecture on the meta device first, swaps in the attention variant you asked for, then loads the state dict to CPU. The clever bit is that loading is lazy: the node just hands you a lazy wrapper, and the actual multi-GB load happens the first time the sampler touches the model. That's why adding the node to a workflow doesn't stall ComfyUI for a minute.

The only output is a MODEL, which wires straight into the pack's TurboDiffusionI2VSampler.

The three inputs you'll actually touch

• model_name - pick from whatever's in diffusion_models/. Set one instance to the high-noise file, another to the low-noise file.
• attention_type - sla (sparse linear attention) is the default and the one that makes the 2-3x speedup happen. original is plain attention for debugging. sagesla needs the optional SpargeAttn package (pip install "spargeattn @ git+https://github.com/thu-ml/SpargeAttn.git"), and honestly, the README warns it may not load on your machine.
• offload_mode - comfy_native (default) lets ComfyUI's async offloading manage VRAM; layerwise_gpu swaps blocks to the GPU just-in-time; cpu_only runs everything on CPU, which is a "does it work at all" mode, not a real option.

attention_type and offload_mode interact, and this is where the pack's own README disagrees with its defaults. The author tested on Windows and reports that only sla + layerwise_gpu was stable on their box; other combos raised load errors or runtime crashes. The shipped default is comfy_native, which is the sensible first try - but if you get weird load failures, switch to layerwise_gpu before you go hunting elsewhere.

Install

ComfyUI Manager: search "TurboDiffusion" (pack title comfyui-turbodiffusion) and install. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/anveshane/Comfyui_turbodiffusion.git
cd Comfyui_turbodiffusion
pip install einops loguru omegaconf pandas

Then restart ComfyUI. The vendored inference code needs omegaconf/loguru/pandas at runtime, and the README's pip line is the version that's actually tested - use it.

Troubleshooting

• "Could not import vendored TurboDiffusion code!" - a known issue on some setups in the early builds. The community workaround was adding an explicit from .. import turbodiffusion_vendor at the top of the try/except in nodes/turbowan_model_loader.py; make sure you're on the latest commit first, then try that.
• OOM on a 5090 - a few users hit CUDA OOM at any resolution, and pinned it to PyTorch 2.9; dropping to PyTorch 2.8 cleared it. If you're on the newest torch and nothing else works, that's the known lever.
• sagesla errors - that's the SpargeAttn dependency, not your fault. sla doesn't need it.
• Slow generations - double-check attention_type didn't fall back to original.