Neutral

This is a LoRA for Krea2 that has been "trained" for exactly 0 steps, on 0 images.

Load it at strength 0.1, at strength 5, at strength -100 if you want, your image will not care. No new style, no extra fingers, no anime filter. It just sits there quietly.

Is it based on anyone else's work? No. Is it useful for making anything? Also no. Is it, in some small way, the purest possible LoRA, an adapter that adapts nothing.

How it was made: all lora_A / lora_B tensors are exact zero matrices, matching the key structure of a real Krea2 LoRA (same shapes, same 28 blocks, same attn/mlp targets). Since the LoRA delta is strength × (B @ A), and both B and A are zero, the output is mathematically zero regardless of the strength value.

Why this exists: I made this to compare downloaded LoRAs against "nothing", a proper baseline.
In my usual comparison workflow, getting a genuine no-LoRA output alongside the LoRA outputs normally means extra routing: switches, bypass nodes, a second branch that skips the loader entirely. Dropping this LoRA into the same loader chain does that job, but this one lets me keep everything in a single unified chain, one loader, one list of LoRAs to test, one weight, no separate no-LoRA path to wire up. Simpler graph, one thing less to maintain.

The honest paragraph: this LoRA is practically neutral, not mathematically perfect down to the last bit. Testing across a strength range produced bit-for-bit identical outputs every time, proof there's no leftover signal in the tensors. But comparing "Neutral active" against "no LoRA node at all" still shows a small, constant, strength-independent difference coming from ComfyUI's model-patching code path itself, not from this file. For reference, a real style LoRA at working strength produces roughly 5x more deviation, affecting most of the pixels.