ADPipelineConfigurator

Every sampler in asyncdiff_comfyui demands an AD_CONFIG input, and this is the node that makes one. It's the configuration heart of the pack - the thing that decides how AsyncDiff slices your diffusion model across your GPUs and how the asynchronous denoising runs. If you're coming from regular ComfyUI where you never think about this stuff, this node is where the pack stops feeling like ComfyUI and starts feeling like a distributed-computing project. Don't panic - four inputs matter, the rest can stay at defaults.

What AsyncDiff actually is (in one paragraph)

It's the NeurIPS 2024 framework by czg1225 and co. that turns the sequential denoising loop into a pipeline. Normally each denoising step finishes before the next starts - a serial chain. AsyncDiff exploits the fact that hidden states barely change between consecutive steps: after a warm-up phase, each component can predict roughly what its input will look like next step, so GPU 2 starts computing while GPU 1 is still finishing. The chain becomes a pipeline, and multiple GPUs work at once instead of one sitting idle. The paper's own demos: ~2.8x faster on SDXL with 4 devices, ~1.8x on AnimateDiff with 2. It's a latency trick, not a quality trick - output isn't bit-identical to a sequential run, and you trade a little fidelity for speed.

The inputs that actually matter

• model_n (default 2, range 2–4) - the important one. The number of pieces the denoiser is split into, which should match the GPUs you're throwing at it. 2 = split in half, 4 = quarters. This is where the speedup comes from, and it's also where people break things: asking for model_n 4 on two GPUs is asking for trouble.
• nproc_per_node (default 2, range 1–4) - processes per node, one per GPU in a single-machine setup. Set it to your GPU count.
• warm_up_steps (default 40) - how many sequential steps before the async pipeline kicks in. More warm-up = output closer to the non-accelerated version, at the cost of speed. The upstream example uses just 1; 40 is a conservative default that prioritizes fidelity.
• time_shift (default off) - leave it off. The upstream docs say only flip this on if the accelerated output comes out with significant noise. If your images look clean, it stays off.
• variant - precision: fp16 default, or bf16 / fp32. fp16 is the right call unless your GPUs choke on it.

The rest are mostly memory and speed toggles: enable_tiling and enable_slicing are diffusers' attention tiling/slicing tricks for fitting bigger workloads into VRAM, xformers_efficient swaps in xformers attention if you have it installed, scale_input/scale_percentage scale the input down (handy when memory is tight), and pipeline_init_timeout is how long to wait for the distributed pipeline to spin up before declaring it dead.

Install

Standard for the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/SlackinJack/asyncdiff_comfyui.git
cd asyncdiff_comfyui
./clone_asyncdiff_repo.sh   # required - pulls the upstream AsyncDiff framework

Restart ComfyUI (or ComfyUI Manager, search "asyncdiff_comfyui"). Heavy dependencies come along for the ride: torch.distributed, diffusers, accelerate, NVIDIA CUDA >= 12.0.

Where people get burned

This node is the source of most "it's broken" reports, and almost all of them are actually multi-GPU setup problems: NCCL failing to initialize, model_n not matching the real GPU count, or nproc_per_node set wrong. Verify with nvidia-smi before blaming the node. And the meta-warning applies to the whole pack: it's multi-GPU or nothing. One card, and all this configuration buys you a slower run than ComfyUI's native sampler. The author's README also flags stable-video-diffusion-img2vid as the currently-supported lane - so if the SD samplers act up, that's a frontier you're walking into, not a bug you caused.