T2IAdapter Custom Weights 🛂🅐🅒🅝

This is the original per-block weights node for T2I-Adapters - four sliders (weight_00 through weight_03) plus a flip_weights toggle. T2I-Adapters are the lightweight alternative to full ControlNets (smaller, faster, the pick when VRAM is tight, per the ControlNet KB), and they only inject at four feature scales, so a T2I-Adapter weights node has exactly four knobs. This is that node in its general-purpose form.

Functionally it matches the ACN_-prefixed T2IAdapter Custom Weights node; the visible difference is the flip_weights option here. Both reweight a T2I-Adapter block by block; use whichever your workflow already references.

What the four blocks do, and the flip

The four blocks roughly span coarse to fine - earlier blocks govern broad structure, later ones finer guidance. All default to 1.0 (uniform). flip_weights (default false) reverses the order, swapping a coarse-biased profile for a fine-biased one without retyping. With only four values it's a small convenience, but it's there.

Realistically, T2I-Adapters are usually a "set overall strength and move on" affair - the whole point of them is being cheap and low-fuss. Per-block weighting is for the occasional case where one scale dominates or the adapter is too weak overall and a single strength dial can't fix it.

Inputs and outputs

• weight_00 … weight_03 (4 required) - per-block T2I-Adapter strengths, default 1.0 each.
• flip_weights (required, default false) - reverse the weight order.
• uncond_multiplier (optional) - 0.0 for "ControlNet is more important" (drops it from the uncond pass), continuous between 0 and 1. Default 1.0.
• cn_extras (optional) - chain in extra weight modifiers.

Outputs: CN_WEIGHTS (into a Timestep Keyframe's cn_weights, or weights_override on Apply Advanced ControlNet) and TK_SHORTCUT (a Timestep Keyframe for a direct timestep_kf connection).

Installing the pack

ComfyUI Manager: Install Custom Nodes → "ComfyUI-Advanced-ControlNet" → restart. Or cd ComfyUI/custom_nodes && git clone https://github.com/Kosinkadink/ComfyUI-Advanced-ControlNet and restart. It's Kosinkadink's pack, pure ComfyUI. No preprocessors (comfyui_controlnet_aux) and no model files bundled - your T2I-Adapter goes in ComfyUI/models/controlnet.

Common gotchas

Four blocks means T2I-Adapters only. Point this at a full ControlNet and the weights won't map - you'll get a compatibility error listing the weight types that ControlNet accepts. Full ControlNets want the [SD1.5] or [Flux] custom-weights node instead.

Second, the T2I-Adapter-vs-ControlNet distinction is the thing to internalize: they load into the same path and feed the same Apply node, but they're different model families with different block counts, and that mismatch is what "the weights node won't fit" almost always means. Finally, if you see both this node and its ACN_-prefixed twin, don't overthink it - they do the same job; grab the one your graph already has, and remember the four values multiply against the Apply node's overall strength.