Nodes/ComfUI-EGAdapterMadAssistant/IPAdapter Mad Assistant V3
ComfyUI Node

IPAdapter Mad Assistant V3

When one layer just needs to dominate

By 11dogzi·Created 2 years ago·Updated 2 years ago· 40
IPAdapter Mad Assistant V3
    • layer_weights
    0layer Weight0.0
    1layer Weight0.0
    2layer Weight0.0
    3layer Weight0.0
    4layer Weight0.0
    5layer Weight0.0
    6layer Weight0.0
    7layer Weight0.0
    8layer Weight0.0
    9layer Weight0.0
    10layer Weight0.0
    11layer Weight0.0
    RandomNo
    seed0

    IPAdapter Mad Assistant V3 widens the twelve layer Weight sliders to -3.0 to 3.0. Beyond the base node's 0..1 and V2's 0..2, you're now solidly in "this layer gets to be the boss" territory.

    Let's be honest about what weights like this mean. The community's standing advice is that a plain IP-Adapter weight around 0.7–0.8 already lets the reference override the prompt. At 3.0 on a layer you're not doing style transfer anymore - you're ordering that layer to reproduce the reference's signal as hard as the model physically can. There's a legitimate use for that: the Mad Scientist approach where you hammer a couple of late style layers hard and keep the early structure layers near zero, so the reference's look comes through without dragging its composition along. But the realistic shape of a V3 workflow is one or two sliders parked at the extremes and the rest barely off zero.

    Everything else is unchanged from the family: twelve sliders in, comma-joined layer:weight string out via the layer_weights STRING output, wired into the IPAdapter Mad Scientist node's layer_weights input (convert that widget to an input first). Random mode fills currently-zero layers from the full -3.0 to 3.0 range - at this spread that's chaos, and it's the mode you probably skip on this node. The seed input is declared but never read, so there's no reproducibility to fall back on either.

    Which one do you want?

    Honestly? Most people never need V3. If you're here it's because you've identified one specific layer that reads too weak at 1.0 or 2.0 and you want to test the extreme before concluding the model just won't do it. The wider ranges in this family aren't progressively better - they're progressively more of the same knob. V3 is where "more range" starts to be a trap rather than a feature.

    Install

    Same as the whole 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 dependencies, no model downloads. You still need ComfyUI_IPAdapter_plus and its model files installed for the Mad Scientist node to exist at all.

    One grounded gotcha: the README describes this pack as "0–10 layers," but the code ships 0–11 (twelve sliders). Don't count layers from the README; count them from the node.

    Categoryipadapter

    Inputs (14)

    NameTypeDefaultDescription
    0layer WeightFLOAT0.0-3–3
    1layer WeightFLOAT0.0-3–3
    2layer WeightFLOAT0.0-3–3
    3layer WeightFLOAT0.0-3–3
    4layer WeightFLOAT0.0-3–3
    5layer WeightFLOAT0.0-3–3
    6layer WeightFLOAT0.0-3–3
    7layer WeightFLOAT0.0-3–3
    8layer WeightFLOAT0.0-3–3
    9layer WeightFLOAT0.0-3–3
    10layer WeightFLOAT0.0-3–3
    11layer WeightFLOAT0.0-3–3
    RandomCOMBONo2 options: Yes, No
    seedINT0-1125899906842624–1125899906842624

    Outputs (1)

    NameTypeDescription
    layer_weightsSTRING