ComfyUI-Krea2GatedLoRAMerge

ComfyUI-Krea2GatedLoRAMerge

Experimental file-level Krea2 LoRA merger with conservative handling for gate layers.

The node writes a new .safetensors LoRA into ComfyUI's output folder. Normal layers can be merged by concatenating LoRA ranks, which exactly represents strength_a * delta_a + strength_b * delta_b, or compressed back to the larger input rank.

Gate layers can use a conflict rule when the two reconstructed deltas point in opposite directions:

• pick_strongest: keep the gate delta with larger norm
• zero_out: drop that gate delta

Gate layers are keys containing .attn.to_gate. / .attn.gate. or .ff.gate. / .mlp.gate..

balance_mode controls whether matched layers are strength-balanced before merging:

• off: use the requested LoRA strengths directly
• per_layer_norm: equalize each pair of matched layer delta norms after the requested strengths are applied

gate_compression controls whether non-conflicting gate merges remain exact rank concatenations or are compressed back down to gate_rank:

• off: exact rank concatenation for non-conflicting gates
• attn_to_gate_only: compress attention gates only
• all_gates: compress attention and MLP gates

Compression is done in low-rank factor space, not by materializing the huge full gate matrix. compression_device=cuda can speed this up if enough VRAM is available.

normal_compression controls ordinary non-gate layers:

• off: exact rank concatenation, larger output
• max_input_rank: compress each merged normal layer back to the larger input rank