Load IPAdapter SD3 Model

If you've ever used an IP-Adapter on SD 1.5 or SDXL, you already know the shape of this thing: load an adapter file, feed a reference image in, and the model steers toward what that image looks like - style, mood, subject - without you training a LoRA. This node is the loader half of that trick for Stable Diffusion 3.5 Large, and it's the most unglamorous node you'll ever wire up. Two inputs, one output, zero creativity. That's the job.

Here's the backstory, because it explains why this pack exists at all. InstantX shipped the official SD3.5 Large IP-Adapter weights in late 2024 but never shipped ComfyUI support - a team member said outright in the release thread that they're a handful of part-timers and "we leave this part to the community." This pack, by longtime community dev Slickytail, is that community part. If you want image conditioning on SD3.5, this is basically the door.

What it actually does

IPAdapterSD3Loader reads your ip-adapter.bin checkpoint from the models/ipadapter folder and builds the two moving pieces the Apply node will later need:

• a TimeResampler - a Perceiver-style resampler (in the Open-Flamingo lineage) that compresses the reference image's SigLIP embeddings into 64 queries. The "Time" part matters: unlike the old Tencent adapters, this one is conditioned on the denoising timestep, which is how the adapter stays aware of where you are in the generation.
• the attention processors that will be patched into SD3.5's MMDiT joint blocks. Everything is cast to fp16 and moved to whatever device you pick.

That's it. You don't tweak anything clever here; the node has exactly two inputs and both are basically choices.

• ipadapter - a dropdown of everything sitting in models/ipadapter. If it's empty, the file isn't there.
• provider - cuda, cpu, or mps. You'll run cuda on a normal rig; mps is the Apple Silicon option; cpu technically works but this is an 8-billion-parameter diffusion model we're talking about, so don't.

The single output, ipadapter (typed IP_ADAPTER_SD3_INSTANTX), plugs straight into ApplyIPAdapterSD3's ipadapter input. That's the whole wiring story.

Installing it

The pack is on the Comfy registry, so ComfyUI Manager finds it by searching ComfyUI-InstantX-IPAdapter-SD3, or you can clone it the old-fashioned way:

cd ComfyUI/custom_nodes
git clone https://github.com/Slickytail/ComfyUI-InstantX-IPAdapter-SD3
# restart ComfyUI

Dependencies are just torch, einops, and diffusers - nothing exotic, and you almost certainly already have all three. The real install is the models:

# the adapter itself → models/ipadapter
# (rename it to something you'll recognize, e.g. ip_sd35l_instantx.bin)
# from https://huggingface.co/InstantX/SD3.5-Large-IP-Adapter

# the vision encoder → models/clip_vision
# siglip_vision_patch14_384.safetensors
# from https://huggingface.co/Comfy-Org/sigclip_vision_384

Gotchas worth knowing

• It's hardcoded for SD3.5 Large. The resampler dimensions are written for Large's architecture, so this will not load for SD3.5 Medium or plain SD3. The README's own TODO says as much - the author wants to infer these from the checkpoint one day.
• The adapter ships as a .bin, not safetensors. It loads through torch.load rather than the safetensors path (.safetensors and .sft are supported too, but the InstantX download is a .bin). Weirdly-named files sometimes trip up security scanners; that's a false alarm here.
• Empty dropdown is the classic failure mode: the file either isn't in models/ipadapter or has an extension the loader doesn't recognize.

It's not a flashy node, but it's the prerequisite for everything the Apply node does - and SD3.5 is niche enough that this is one of the only games in town.