IPAdapter Mad Assistant V2

IPAdapter Mad Assistant V2 is the same slider-to-string helper as the base node, but the twelve layer Weight sliders span -2.0 to 2.0. That's the whole difference - and it's a useful one if you keep maxing out the base node's 1.0 ceiling on your style layers.

Here's the situation where this pays for itself: you've set a high-layer weight to 1.0, the reference is doing most of what you want, and you still want more of it. On the base node you're stuck - 1.0 is the top. V2 gives you 2.0 of headroom, plus the negatives to subtract a layer you're fighting. Weights above 1.0 push the reference harder on that layer; values toward -2.0 suppress it. The one caution is that you're operating past the weight where the reference starts overriding the prompt, so raise the hot layers and keep the early structure layers low or you'll lose your composition.

Mechanically it's identical to the base node: twelve sliders in, a comma-joined layer:weight string out via the layer_weights STRING output, which you wire into the IPAdapter Mad Scientist node's layer_weights input (right-click → Convert to Input first). Random mode behaves the same too, except zero layers get filled from the full -2.0 to 2.0 range - noticeably wilder than the base's 0..1. Pin any layer you care about to a non-zero value if you run random, and note the seed input is declared but never read, so randomness isn't reproducible.

Which one do you want?

Reach for V2 when you've already tried the base node and the answer was "I need more than one." If you're just getting into per-layer IPAdapter tuning, start on the base node - the extra range here mostly means extra ways to blow out an image. V2 is the "I know exactly which layer is too weak" upgrade.

Install

Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/11dogzi/ComfUI-EGAdapterMadAssistant

Restart ComfyUI, or search ComfUI-EGAdapterMadAssistant in ComfyUI Manager. No pip, no requirements, no model downloads - pure standard library. You still need ComfyUI_IPAdapter_plus and its models for the output to matter.

And while you're here: the README calls V2 the "-1 to 1" node. The shipped code disagrees - V1 is the -1..1 one, V2 is -2..2. When the README and the code disagree, the code wins.