IPAdapter Advanced V2

This is the workhorse. IPAdapter Advanced V2 is the version of image prompting you'll reach for when you want style, mood, or a vague "looks like this" out of a reference image - and you want control over how hard that reference pulls. It's the middle of the IP-Adapter family: more knobs than the Simple node, fewer than the Mad Scientist. If you're building a workflow that carries a visual identity from one generation to the next, this is the node you'll actually keep.

How it works

IP-Adapter was Tencent's August 2023 trick for making image prompting cheap: instead of training a LoRA to teach a model what something looks like, you hand it a reference and a ~22M-parameter adapter injects the image's CLIP embedding through its own cross-attention path. The text path stays frozen, so your prompt keeps working next to the image conditioning. This V2 family is the rewritten API that cubiq shipped in 2024 - and this pack is a renamed copy of that codebase so it can coexist with the older v1-style nodes without name collisions.

The inputs that matter

• weight (default 1.0, range −1 to 5) - how strongly the reference influences the output. This is your main dial. Community consensus: 0.3–0.6 is the sweet spot; past about 0.7–0.8 the reference overrides the prompt.
• start_at / end_at (0 to 1) - which slice of the denoising steps the adapter is active. Running it across all steps measurably softens detail; running it only for the early steps preserves it.
• weight_type - the curve of that strength. linear for a flat pull; ease out if you want the reference strong early and the prompt to take over later; style transfer and composition are presets that bias the influence toward look versus layout.
• combine_embeds - how multiple reference images merge (concat, add, subtract, average, norm average). concat keeps them as separate tokens; average blends toward a middle look.
• embeds_scaling - the least-understood one: how the image embedding is applied to the attention K/V. V only is the default and fine for nearly everyone; the K+V options change behavior for specific weight types and are worth leaving alone until something's not working.

Optional inputs do what you'd guess: image_negative is a "don't look like this" reference (opposite of a negative prompt - you supply an actual image), attn_mask restricts the conditioning to a region, and clip_vision lets you supply a CLIP vision model if you didn't use a unified loader. The single output is MODEL, which feeds straight into your KSampler.

Installing it

ComfyUI Manager: search "ComfyUI_IPAdapter_plus_V2". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_IPAdapter_plus_V2

Restart. The adapter weights go in ComfyUI/models/ipadapter (e.g. ip-adapter-plus_sd15.safetensors or the SDXL equivalents), and the CLIP vision model in ComfyUI/models/clip_vision.

Where people get burned

The adapter needs both a model and CLIP vision loaded, usually via IPAdapter Unified Loader V2 or IPAdapter Model Loader V2 plus a CLIP vision loader - wire ipadapter and clip_vision into the inputs or you'll get a "Missing CLIPVision" error. And remember this pack is a snapshot of cubiq's V2 codebase: it doesn't pick up upstream fixes, so if you update the original pack, these nodes won't move with it.