Nodes/ComfUI-EGAdapterMadAssistant/IPAdapter Mad Assistant V1
ComfyUI Node

IPAdapter Mad Assistant V1

Negative weights, for dialing a layer back and then some

By 11dogzi·Created 2 years ago·Updated 2 years ago· 40
IPAdapter Mad Assistant V1
    • 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 V1 is the same slider-to-string helper as the base node, with one difference that changes what it's for: the twelve layer Weight sliders run -1.0 to 1.0 instead of 0.0 to 1.0. You get the ability to go negative on any of the 0–11 layers.

    Why would you want negative? Setting a layer to -1.0 doesn't just turn the reference image off on that layer - it pushes against it, actively subtracting the image embedding's contribution. The practical use is when a style transfer keeps dragging the composition or when one specific layer reads too hot and you want to suppress it rather than simply zero it. The community mostly lives in the 0..1 range; negative values are the "turn this off, and then some" move. If a generation is over-lifting structure, dropping an early layer into the negatives is often what actually fixes it.

    Everything else is the base node. It builds the layer:weight comma string (output layer_weights, STRING), you convert Mad Scientist's layer_weights widget to an input and wire it up, and Random mode works the same way - except the random fill now draws from -1.0 to 1.0 instead of 0 to 1, so random runs can quietly zero-invert a layer. The seed input is declared but never used in the code, so random mode stays unreproducible.

    Which one do you want?

    Honest answer: the base node, 90% of the time. V1 earns its keep when you've identified a layer that's too present - composition fighting you, or one style band dominating. Keep your daily workflows on the 0..1 node and drop in V1 when you need the subtraction.

    One README trap worth knowing: the pack's README credits V2 as the "-1 to 1" variant. It isn't - per the shipped code V1 is -1..1 and V2 goes to -2..2. If a workflow you downloaded mentions the "gentle negative" version by the wrong number, this is the one.

    Install

    Shared with the whole pack - it's the same single-file repo:

    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 models for anything to render.

    If the output does nothing, check that you converted Mad Scientist's layer_weights widget to an input before wiring it - that's the failure mode that looks exactly like a broken node.

    Categoryipadapter

    Inputs (14)

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

    Outputs (1)

    NameTypeDescription
    layer_weightsSTRING