T2IAdapter Soft Weights 🛂🅐🅒🅝

Same idea as Scaled Soft Weights, but shaped for T2I-Adapters instead of ControlNets. T2I-Adapters are the lightweight cousins of ControlNet - smaller, faster, easier on VRAM, and the go-to when VRAM is your constraint on SDXL. The catch is they inject at only four points in the network, not the dozen-plus a ControlNet uses, which is why this node gives you exactly four weight sliders instead of a single softness multiplier. You're setting the strength at each of the adapter's four blocks directly.

If you're trying to soften a T2I-Adapter's grip so your prompt has more room, this is the correct node - feeding a T2I-Adapter into the ControlNet-shaped Scaled Soft Weights won't match its structure.

How it works

A T2I-Adapter conditions the base model at four feature levels, roughly coarse-to-fine. This node lets you weight each level on its own. The defaults are already a gentle ramp - the earliest (coarsest) block is dialed down and each subsequent block leans in harder - which softens the broad conditioning while keeping detail-level guidance intact. Raise the early sliders for a firmer overall hold; lower the later ones to let the model take back fine detail. As with the ControlNet version, there's a separate uncond_multiplier for the effect on the unconditional pass, so you can reproduce the "control is more important" behavior by pushing it to 0.

The inputs and outputs that matter

Four required sliders, one per adapter block:

• input_0 (default 0.25), input_1 (0.62), input_2 (0.825), input_3 (1.0) - coarse to fine. The defaults are a sensible soft ramp; treat them as a starting point.

Optional uncond_multiplier (default 1.0) controls influence on the negative pass - set to 0 for the strongest, most control-dominant behavior. Optional cn_extras carries extra weight data if you're combining with something like Anima's mask extras.

Outputs are CN_WEIGHTS - into weights_override on the Apply Advanced ControlNet node, or cn_weights on a Timestep Keyframe - and TK_SHORTCUT, a pre-built Timestep Keyframe wrapping these weights so you can plug straight into a loader's timestep input.

How to install it

ComfyUI Manager: search ComfyUI-Advanced-ControlNet, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Kosinkadink/ComfyUI-Advanced-ControlNet

then restart. No dependencies. You'll need a T2I-Adapter model in ComfyUI/models/controlnet for this to apply to - the node makes the weights, not the adapter.

Common issues & troubleshooting

"Incompatible weight type." You've likely fed these weights into a standard ControlNet, or fed ControlNet weights into a T2I-Adapter. The two have different block counts and aren't interchangeable - use this node only with T2I-Adapters, and Scaled Soft Weights with ControlNets. The error will tell you which types are valid.

The four sliders feel opaque. They're coarse-to-fine, so input_0 affects broad layout most and input_3 affects fine detail most. If the adapter is overpowering the composition, pull input_0 down; if detail is being ignored, raise the later ones.

You just want one softness dial. T2I-Adapters don't work that way - the four points are the control surface. Adjust them together if you want a uniform change.