IPAdapter

This is the front door. IPAdapter is the basic apply node - the stripped-down version of IPAdapter Advanced with the intimidating settings removed. If you've never used IP-Adapter before and you want to see a reference image bleed into your generation without reading a settings manual, start here.

The idea, in one line: IP-Adapter is a one-image LoRA. You give it a reference picture, it encodes that picture with CLIP vision, and it injects the result into your model through a dedicated cross-attention path that runs alongside your text prompt. Your prompt keeps working; the image conditioning rides in next to it. Style, subject, mood - whatever's in the reference starts showing up in the output, no training required.

Inputs and outputs

Three connections make it go: model (your checkpoint), ipadapter (from a loader - the Unified Loader is easiest), and image (your reference). It returns a patched MODEL that plugs straight into your KSampler. Minimal graph: checkpoint → Unified Loader → IPAdapter → KSampler.

The handful of settings:

• weight - how strongly the reference pushes, default 1.0 (range −1 to 3). The single most useful thing to know: turn this down. Around 0.8 is a good starting point, and 0.3–0.5 if you want your text prompt to stay in the driver's seat. Crank it past ~0.8 and the reference takes over, prompt coherence be damned.
• weight_type - here it's the friendly three-option version: standard (balanced), prompt is more important (leans back toward your text), and style transfer (lifts the look, less the subject). That's it - no fifteen-way curve menu like the Advanced node.
• start_at / end_at - the slice of the sampling run the adapter is active for, as fractions from 0 to 1. Leave them at 0 and 1 to start.

The one optional input is attn_mask - a mask to confine the effect to part of the frame. Everything else the Advanced node exposes (combine_embeds, embeds_scaling, negative image) is intentionally hidden here to keep it approachable.

When to graduate

This node is a great teacher and a fine daily driver for simple single-reference style transfer. You'll want to move up to IPAdapter Advanced when you need any of: multiple reference images blended together, the InstantStyle-style precise style transfer/composition weight types, a negative-image input, or the deeper attention-scaling controls. There's no penalty for starting here and switching later - it's the same underlying adapter, just more knobs.

Installing the pack

ComfyUI Manager: open the Custom Nodes Manager, search "IPAdapter plus", install, restart. Manual route: cd ComfyUI/custom_nodes && git clone https://github.com/cubiq/ComfyUI_IPAdapter_plus, then restart ComfyUI. The pack is small; the real downloads are the models - a CLIP vision encoder in ComfyUI/models/clip_vision (ViT-H for SD1.5, ViT-bigG for SDXL) and an IP-Adapter model in ComfyUI/models/ipadapter. If the node errors on load, update ComfyUI - the pack (built by Matteo, aka cubiq, and something of a community favorite) always wants the latest core.

Reality check on scope: this is an SD1.5 and SDXL tool, strongest on SD1.5, and it does not load on Flux. If you're on a newer architecture, this whole pack isn't your answer - but for classic checkpoints, IP-Adapter is still the cleanest way to prompt with a picture.