K2 Projector

K2 Projector is the K2 Lab suite's control for one of the least-known but strongest levers in Krea 2: the txtfusion.projector. That's the Linear(12 → 1) that mixes Krea 2's twelve Qwen3-VL hidden-state layer taps into the final text vector. Shift that mix and you change which language level dominates the image - the whole-model semantic flavor. It's the mechanism behind the well-known "projector LoRAs" that circulate for Krea 2, and this node applies the same delta without touching base weights.

In plain terms: the model's style, aesthetic direction, and "how literally it reads the prompt" are shaped by which of those twelve layers it leans on. This node moves the mix. If you've seen people talk about "filter bypass" or "style shift" LoRAs for Krea 2, this is the knob those LoRAs were turning.

The inputs

• preset - built-in vectors from the published K2Lab reference table: none (all zeros), filter_bypass2, filter_bypass3, skc3vo, z0jglf, and custom. skc3vo and z0jglf are the two big named projector LoRA styles, here as approximations.
• multiplier - signed scale on the whole vector. 0 = no effect, negative reverses it. This is the strength dial.
• identity_protection - 0..1. At 1.0, identity prompt tokens stay on the unmodified baseline mix, so a strong style shift doesn't deform faces. Leave it at 1 unless you specifically want the style to touch identity.
• enabled - off passes the model through untouched.
• custom_values - twelve numbers, comma or space separated, only used with preset custom. Handy if you have a projector vector from somewhere else.

Outputs are the projector object (wire into K2 Compose's projector input) and values, the current vector as text.

The honest caveat

The built-in presets mirror the published K2Lab reference table, and the pack's own documentation is upfront that the table is an approximation - for skc3vo and z0jglf, several signs differ from the values in the actual LoRAs. So if you have the actual projector LoRA file, use K2 Projector from LoRA instead, which extracts the exact twelve values straight from the file. The presets are the "I don't have the file, just give me the vibe" path; the LoRA node is the precise path.

Install

Same as the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/ping1979ping/comfyui-FVMtools

Restart ComfyUI, find it under FVM Tools/K2. No extra dependencies beyond the standard Krea 2 stack.

Worth knowing before you lean on it: this is the single strongest semantic lever in the model, so start at multiplier 0.5, not 2. A projector delta at full strength is a style change you can't un-see, and pairing it with identity_protection off is how people end up with distorted faces and blame the model.