ADModelLoader

ADModelLoader is the least glamorous node in the pack and the one you can't skip. Every sampler in asyncdiff_comfyui takes an AD_MODEL input, and this is where that object comes from. Think of it as the checkpoint loader for a pack that never touches ComfyUI's normal model loading.

There's a reason for that, and it's the thing to understand about this whole pack before you wire anything. AsyncDiff works on diffusers pipelines, not on ComfyUI's internal model format. When you use the built-in CheckpointLoaderSimple, you get a model that ComfyUI's own KSampler understands. None of the AD samplers here can eat that. They want a model in a diffusers-style layout - the folder-of-files arrangement (model_index.json, unet/, vae/, text_encoder/) that HuggingFace checkpoints ship in - so ADModelLoader is how you hand one over.

What you actually do with it

It's a single required input, model, which is a dropdown populated with whatever the pack can see. You pick a model, you get an AD_MODEL out, and you run that into the model input on whatever sampler you're using - ADSDSampler for SD 1.5/2.1/SDXL/SD3 text-to-image, ADSVDSampler for Stable Video Diffusion, ADSDUpscaleSampler for the x4 upscaler.

Two honest caveats, because this node is where beginners trip over both:

• Where does it look for models? The pack's README doesn't document the folder it scans, and that's a real gap. After you install, open the pack's directory and look for a models folder - drop your diffusers-format checkpoints there and refresh the dropdown. If it stays empty, that's the first thing to check.
• The README's only support note says stable-video-diffusion-img2vid. The SD sampler nodes exist and map to what the upstream AsyncDiff repo supports, but the author's stated "currently supports" line is SVD. If you're here for video, you're on the well-trodden path. If you're here for SD images, you may be the test pilot.

Install

ComfyUI Manager (search "asyncdiff_comfyui"), or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/SlackinJack/asyncdiff_comfyui.git

Then run the step the README insists on before first use:

cd ComfyUI/custom_nodes/asyncdiff_comfyui
./clone_asyncdiff_repo.sh

That script pulls in the upstream AsyncDiff framework (czg1225/AsyncDiff) that every node in the pack calls. Skip it and nothing works. Then restart ComfyUI.

The real requirement nobody says out loud

This pack is multi-GPU or nothing. AsyncDiff is model parallelism - it splits the denoiser across devices, so the entire point collapses on a single card. The upstream repo wants NVIDIA GPUs, CUDA >= 12.0, PyTorch >= 2.0.1, plus torch.distributed, diffusers, and accelerate. If you're on one GPU, ADModelLoader will happily load a model and then the sampler will just be slower than ComfyUI's native path. Know that before you invest an evening.

Output is the single AD_MODEL wire. That's the whole node - a dropdown and a wire, but it's the one that gets the whole graph on its feet.