Krea Qwen3-VL Layer Bands

The layer-weight node in this pack gives you a smooth ramp from "texture" to "semantics," but a ramp is a blunt instrument. If you actually want to know which Qwen3-VL layers carry structure, which carry texture, and which carry meaning in Krea 2, you need to poke bands independently. That's exactly what Krea Qwen3-VL Layer Bands is for: it splits the 12-layer conditioning stack into four bands of three and scales each one on its own.

The four bands

The inputs, per the tooltips:

• low - layers 0–2, the lowest-level texture / colour.
• mid_lo - layers 3–5.
• mid_hi - layers 6–8.
• high - layers 9–11, semantics / concept.

Each defaults to 1.0 (no-op), runs 0–3. The mechanic is a plain conditioning transform - reshape the fused tensor, apply one scale per band, re-flatten. No model patching, and like its sibling it keys off Krea 2's 12×2560 width, so any other conditioning passes through untouched. All four at 1.0 is an exact no-op.

The intended use is diagnostic

Turn one band up and the rest to ~1 to isolate what it contributes, and build a mental map of your model. "Raise high and the subject's identity locks in harder" is the kind of thing you only learn by feeling it out, and this is the tool that lets you. The source even lists which actual transformer taps each band covers (low = taps 2, 5, 8; high = taps 29, 32, 35), which tells you something about how spread out Krea's conditioning taps are.

Honest expectations: this is a probe more than a final-workflow node. It's rare that you leave all four bands at weird values in a shipped workflow - but it's extremely common to spend an evening with this node, find that your model carries texture in band 1 and identity in band 4, and then go implement that insight with the simpler Layer Weight ramp. That's a completely normal way to use an experimental node, and it's clearly what the author built it for.

Where people get burned: band semantics aren't universal. They're the author's observed mapping (low = texture, high = concept), and while it matches the usual transformer picture, your checkpoint's aggregation can shift things. So don't assume high is always semantics - verify it on your model, which is the whole point of the node anyway.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzExperimental

or search "Nynxz" in ComfyUI Manager, then restart. Requires Krea 2 loaded properly (checkpoint + Qwen3-VL encoder + Qwen-Image VAE), and a current enough ComfyUI for the V3 node API.