Nodes/ComfyUI_IPAdapter_plus_V2/IPAdapter Mad Scientist V2
ComfyUI Node

IPAdapter Mad Scientist V2

Every knob, all at once

By chflame163·Created 2 years ago·Updated 2 years ago· 55
IPAdapter Mad Scientist V2
  • model
  • ipadapter
  • image
  • image_negative
  • attn_mask
  • clip_vision
  • insightface
  • MODEL
weight1.00
weight_faceidv21.00
weight_type
combine_embeds
start_at0.000
end_at1.000
embeds_scaling
layer_weights

IPAdapter Mad Scientist V2 is the apply node for people who have run out of other knobs. It's the Advanced node plus FaceID capability plus per-layer weight control, all in one panel: weight, weight_faceidv2, weight_type, combine_embeds, start_at/end_at, embeds_scaling, optional image_negative, attn_mask, clip_vision, and insightface - and the layer_weights string that no other apply node offers. It's the "yes, and" node of the pack. It's also, predictably, a node you should graduate into rather than start on.

What layer_weights actually buys you

The headline feature is layer_weights - a string like 0:1.0, 1:0.5, 2:0.0, ... that scales the adapter's influence per attention layer across the model (12 layers in the standard structure). Early layers lean toward composition and layout; later layers carry texture and fine detail. So a layer-weights profile is how you say "let the reference control the scene structure but stop it from stamping its texture onto my subject," or the reverse. That's real power and it's genuinely opaque, which is why this pack ships IPAdapter Layer Weights Slider - twelve sliders that generate the string for you. If you have the slider, use it; typing these strings by hand is how you earn the Mad Scientist name.

How it behaves as a FaceID node

Because it carries the FaceID machinery (insightface input, weight_faceidv2 up to 5), it can do everything the dedicated FaceID nodes do - but the Mad Scientist is really where you land when you're combining identity conditioning with layer sculpting. One node, both capabilities. The face_image debug output the FaceID nodes have isn't here, so you lose that "did it find the right face" check; for pure face work the dedicated node is still the better tool.

When to reach for it (and when not to)

Reach for it when you've dialed weight and start/end as far as they'll go and you need per-layer control, or when you want identity plus layer weights in a single patch. Don't reach for it first - everything on it that isn't layer_weights is the same as Advanced or FaceID, and the extra surface is just more to misread. The pragmatic ladder: Advanced → FaceID for faces → Mad Scientist when layer weights are the point.

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. FaceID usage additionally needs pip install insightface and its models in ComfyUI/models/insightface.

Gotchas

Layer weights interact with the normal weight by multiplication - a layer weight of 0 kills that layer regardless of the main weight, and a blank layer_weights string means "no per-layer override," which is the safe default. And the standing licensing note: anything built on FaceID (InsightFace) embeddings is non-commercial. This pack is a renamed snapshot of cubiq's V2 codebase, so it won't track upstream fixes on its own.

Categoryipadapter/dev

Inputs (15)

NameTypeDefaultDescription
modelMODEL
ipadapterIPADAPTER
imageIMAGE
weightFLOAT1.00-1–5
weight_faceidv2FLOAT1.00-1–5
weight_typeCOMBO15 options: linear, ease in, ease out, ease in-out, reverse in-out, weak input, +9
combine_embedsCOMBO5 options: concat, add, subtract, average, norm average
start_atFLOAT0.0000–1
end_atFLOAT1.0000–1
embeds_scalingCOMBO4 options: V only, K+V, K+V w/ C penalty, K+mean(V) w/ C penalty
layer_weightsSTRING
image_negativeoptIMAGE
attn_maskoptMASK
clip_visionoptCLIP_VISION
insightfaceoptINSIGHTFACE

Outputs (1)

NameTypeDescription
MODELMODEL