ADIPAdapterLoader

ADIPAdapterLoader is how you get IP-Adapter's image prompting into the AsyncDiff pack. One reference picture in, style and identity out - no LoRA training required. It's the same trick that made IP-Adapter the default no-training identity tool of the SD 1.5 era, repackaged for a pack that runs its own diffusers pipelines.

The reason this node exists is the same reason ADControlNetLoader does: asyncdiff_comfyui doesn't speak ComfyUI's native wire formats. IP-Adapter models from ComfyUI_IPAdapter_plus or the core pack produce outputs those samplers can't read. This loader produces an AD_IPADAPTER that the pack's AnimateDiff sampler actually accepts.

What it does

IP-Adapter works by injecting an image embedding through its own cross-attention path, separate from the text path. That decoupling is the whole trick - the reference image guides style, identity, or composition while your text prompt still works normally, all from a ~100MB adapter file instead of a fine-tune. Want a character to keep their face across an animated clip? Reference frame in, done. Want a consistent art style across the whole video? Same input.

The inputs are minimal:

• ip_adapter - dropdown of the IP-Adapter models the pack can see.
• scale - default 1.0. How strongly the reference image asserts itself. This is the knob you'll actually turn: 0.5–0.8 keeps the reference as a flavor, 1.0+ pushes the image to dominate.

Output is a single AD_IPADAPTER wire, and it goes to exactly one place: the optional ip_adapter input on ADADSampler. It pairs naturally with the pack's image input on that sampler - reference frame in, IP-Adapter steering it, AsyncDiff splitting the compute across your GPUs.

Install

Standard for this pack:

cd ComfyUI/custom_nodes
git clone https://github.com/SlackinJack/asyncdiff_comfyui.git
cd asyncdiff_comfyui
./clone_asyncdiff_repo.sh   # pulls the upstream AsyncDiff framework - don't skip it

Restart ComfyUI. (Or ComfyUI Manager, search "asyncdiff_comfyui".)

Where people get burned

The pack is multi-GPU by design - one card and this whole exercise is pointless. Also remember the ecosystem reality: IP-Adapter's upstream development stopped in early 2024 and nothing in the family runs on Flux-era architectures, so you're firmly in SD 1.5/SDXL territory here, which is exactly where this pack's AnimateDiff and SVD lanes live. And if your dropdown is empty, it's the model-folder problem again - the README doesn't say where the pack scans, so peek inside the pack directory after install.

One honest note: the README's supported path is stable-video-diffusion-img2vid, and IP-Adapter feeds the AnimateDiff sampler that's outside that tested lane. The tooling is real and it works, but you're the first one through the door.