Skip to content

lora-merge-strategies.md

Latest commit

 

History

History
39 lines (29 loc) · 2.15 KB

lora-merge-strategies.md

File metadata and controls

39 lines (29 loc) · 2.15 KB
Raw

LoRA Merge Strategies

Back to README

Each architecture applies its own filter-based preset to tensor categories:

LTX-2.3 types (All, Video, Audio):

  • Classifies tensors into: attn_qkv, attn_out, ff_in, ff_out, audio_attn, audio_attn_out, audio_to_video_attn, video_to_audio_attn, audio_ff_in, audio_ff_out, caption_projection, patchify_or_output, norm, other
  • Preserves structural layers (adaln, gate logits, baked VAE/text/audio modules)
  • All merges all non-preserved weights (normal ComfyUI LoRA load behavior)
  • Video merges only weights without audio in their key. Audio merges every weight with audio in its key, including cross-modal bridge weights

WAN 2.2 strategies (Balanced, Motion, Visuals):

  • Classifies tensors into: self_attn_qkv, self_attn_out, cross_attn_qkv, cross_attn_out, ffn_in, ffn_out, modulation, caption_projection, patchify_or_output, norm, other
  • No Audio strategy (WAN 2.2 has no audio components)
  • Preserves modulation.lin, patch_embedding, and baked companion modules
  • Norm layers always get 0.0 multiplier (untouched)

Krea 2 strategies (Balanced, Style, Content, Detail):

  • Classifies tensors into: attn_qkv, attn_out, attn_gate, ff_in, ff_out, text_fusion, structural, other
  • Style boosts attention (qkv/out/gate), reduces text_fusion — for aesthetic/style LoRAs
  • Content boosts feed-forward, moderates attention — for subject/content LoRAs
  • Detail applies mild global boost — for quality/detail LoRAs
  • Preserves modulation.lin, tproj, tmlp, txtmlp, first/last layers, txtfusion.projector, norm.scale, qknorm

Strength limit: effective strength (global x per_lora) capped at +/-3.0 to prevent black images on Krea 2 gate tensors. Merge device: CPU/CUDA/auto with VRAM headroom check and OOM fallback.

Supported LoRA Formats

Standard .safetensors LoRAs and ComfyUI .diff format are both supported.

Recipe Reload

Every quantization and LoRA merge writes a human-readable .txt recipe alongside the output. Click Load Recipe in the UI to reload a previous run's exact settings (source, output name, all LoRAs, formats, strategy, strength) — useful for reproducing results or iterating on parameters.