Krea 2 (and other gated DiTs) put multiplicative sigmoid gate projections
(attn.gate / mlp.gate) in every block - ~25% of a Krea 2 LoRA. Their edits
have outsized, compounding effect (stacked LoRAs fighting -> deformities)
but are modest in norm, so tame_layers misses them. gate_strength targets
them by name and scales just those layers. Wired into both nodes; non-gate
layers left untouched. Bump 1.2.0.
After keep_energy truncates a layer, optionally rescale the kept singular
values so the layer's nuclear norm matches the original — removes the
conflict-prone tail without weakening the LoRA's effect. Off by default.
Drop-in LoRA loader with optional per-layer SVD energy trimming (kohya
resize math) and hot-layer taming to reduce artifacts from stacked LoRAs
and turbo/distilled models. Passes through DoRA/tucker/unknown keys and
hands the cleaned state dict to ComfyUI's standard load_lora_for_models.
CoreyCorza