Krea Qwen3-VL Layer Weight

Krea 2 doesn't condition on a single Qwen3-VL layer - it conditions on a stack of 12 hidden-state taps, kept as one long conditioning tensor, and learns to aggregate them. And here's the part that makes this node interesting: early taps carry low-level/textural signal, late taps carry semantics. Low layers are how it looks; high layers are what it is. Krea Qwen3-VL Layer Weight lets you rebalance that stack before it ever reaches the model, biasing the output toward detail or toward concept.

The gentlest lever

It's the gentlest lever this pack offers, and that's its selling point. The mechanism is a pure conditioning-tensor transform: reshape the fused 12×2560 tensor back into its 12 layers, apply a linear ramp that starts at low_level (weight on the early/texture layers) and ends at semantic (weight on the late/concept layers), and re-flatten. No model patch, no attention surgery, no guidance hacking - which is why the source notes it's the lever most likely to survive on Turbo, where heavier surgery tends to break.

The inputs are exactly three: conditioning plus the two ramp endpoints. At low_level = 1.0 and semantic = 1.0 the ramp is flat and the node is a no-op. Push semantic above 1 and late layers get louder - stronger adherence to what the thing is, at the cost of the literal look. Push low_level up and detail/texture wins, and semantic can drift. The tooltips are worth quoting because they set expectations right: ">1 pushes detail and literal features; <1 lets semantics dominate."

The guardrails

Two guardrails matter. First, it's Krea 2 only - it keys off the 12×2560 = 30720 conditioning width, and any other conditioning passes through untouched, so you can drop it on a wire without fear of breaking a non-Krea model. Second, stay moderate: reweighting still feeds Krea 2's learned aggregator, and extreme values push it out of distribution. There's no clamp beyond 0–3, so "3 on everything" will look like a model having a bad day, not a better image.

It's flagged experimental - this pack wears that word honestly - and it's the more surgical cousin of Layer Bands (which splits the stack into four independently-scaled bands for hunting down where effects live) and Layer Mix (which splices two whole conditionings together for style transfer). If you just want "more detail, less concept drift," start here with something like low_level 1.2 / semantic 0.9 and see what moves.

Installing it

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

or search "Nynxz" in ComfyUI Manager, then restart. No dependencies beyond Krea 2 itself - remember that's the checkpoint plus the ~8GB Qwen3-VL text encoder plus the Qwen-Image VAE; people routinely forget the encoder and then blame the node.