ethanfel-ComfyUI-LoRA-Optimizer/README.md

---

A ComfyUI node that automatically analyzes your LoRA stack and selects the best merge strategy per weight group — diff-based merging, TIES conflict resolution, DARE/DELLA sparsification, per-prefix adaptive decisions, SVD patch compression, architecture-aware key normalization, and auto-tuned parameters. Two nodes: LoRA Stack (build input) and LoRA Optimizer (analyze + merge).

The Problem

---

Should I Merge This LoRA?

---

Nodes

LoRA Stack

Builds a list of LoRAs for the optimizer. Chain multiple Stack nodes to add any number of LoRAs.

Inputs: LoRA selector, strength, optional previous LORA_STACK

Outputs: LORA_STACK

---

LoRA Stack (Dynamic)

Single node with adjustable slot count (1–10) — replaces chaining multiple Stack nodes.

Mode	Behavior
Simple	One strength slider per LoRA (applies to both model and CLIP)
Advanced	Separate model_strength and clip_strength per LoRA for independent control

Accepts an optional lora_stack input to chain with other Stack nodes.

Outputs: LORA_STACK

---

LoRA Optimizer

The auto-optimizer. Takes a LORA_STACK, analyzes the LoRAs, and automatically selects the best merge mode and parameters per weight group. Outputs the merged result plus a detailed analysis report with a block strategy map.

Also accepts standard tuple-format stacks (lora_name, model_strength, clip_strength) from Efficiency Nodes, Comfyroll, and similar packs.

Uses a two-pass streaming architecture for low memory usage:

• Pass 1 (Analysis): Computes weight diffs per prefix, samples conflict and magnitude statistics per prefix, then discards the diffs. Only lightweight scalars are kept.
• Pass 2 (Merge): Recomputes diffs per prefix, looks up that prefix's conflict data, picks the optimal strategy for it, and merges. Each prefix is freed after merging. Non-TIES patches are SVD-compressed to low-rank by default.

Peak memory is ~one prefix at a time (~260MB) regardless of LoRA count or model size. GPU-accelerated on both passes.

What It Analyzes

• Per-LoRA metrics (rank, key count, effective L2 norms)
• Pairwise sign conflict ratios per prefix (sampled for efficiency)
• Pairwise cosine similarity (directional alignment between LoRAs)
• Magnitude distribution per prefix
• Key overlap between LoRAs

Per-Prefix Adaptive Merge

The key insight: two LoRAs may overlap in some model blocks but not others. A face LoRA and a style LoRA might only conflict in attention layers 4-7, while the rest of the model is touched by only one of them.

Instead of picking one global strategy (which either wastes TIES trimming on non-overlapping blocks or misses real conflicts), the optimizer decides per weight prefix:

Condition	Strategy
Only 1 LoRA touches this prefix	weighted_sum — full strength, no dilution
2+ LoRAs, sign conflict <= 25%	weighted_average — compatible, simple merge
2+ LoRAs, sign conflict > 25%	ties — resolve conflicts with trim/elect/merge
Magnitude ratio > 2x at prefix	total sign method (stronger LoRA dominates)
Magnitude ratio <= 2x at prefix	frequency sign method (equal votes)

This means non-overlapping regions keep 100% of their LoRA's effect, while genuinely conflicting regions get proper TIES resolution.

TIES Merging

The optimizer automatically selects TIES-Merging (Trim, Elect Sign, Disjoint Merge — Yadav et al., NeurIPS 2023) on prefixes where sign conflicts are detected between LoRAs.

DARE / DELLA Sparsification

DARE and DELLA sparsify each LoRA's diff before merging, reducing parameter interference between LoRAs. Available in two modes: standard (drops weights everywhere) and conflict-aware (only drops weights where LoRAs actually interfere).

Method	How It Works
DARE	Bernoulli random mask at given density. Survivors rescaled by 1/density to preserve expected value. Fast and unbiased.
DELLA	Per-row magnitude ranking. Low-magnitude elements get higher drop probability, high-magnitude elements are kept. More surgical than DARE.
DARE (conflict-aware)	Same as DARE, but only applied at positions where 2+ LoRAs push in opposite directions. Same-sign positions (where LoRAs reinforce each other) are left untouched.
DELLA (conflict-aware)	Same as DELLA, but only at conflict positions. Unique contributions from each LoRA are fully preserved.

Why conflict-aware? Standard sparsification drops weights everywhere — including positions where only one LoRA contributes, or where multiple LoRAs agree. This destroys useful signal. Conflict-aware variants compute a sign-conflict mask first: positions where LoRAs push in opposite directions (actual interference). Only those positions get sparsified. The result: interference is reduced without sacrificing unique features.

Interaction with merge strategies:

• TIES mode: DARE/DELLA replaces the TIES trim step (both achieve sparsification, no need for both)
• Other modes: Applied as preprocessing before the merge operation

Setting	Default	Options
sparsification	disabled	disabled, dare, della, dare_conflict, della_conflict
sparsification_density	0.7	Fraction of parameters to keep (lower = more aggressive)

Auto-Strength

When auto_strength is set to enabled, the optimizer automatically reduces per-LoRA strengths before merging to prevent overexposure from stacking. This is especially useful on distilled/turbo models where 2+ LoRAs at full strength cause blown-out results even with optimal merge mode selection.

The algorithm uses interference-aware energy normalization: it measures pairwise cosine similarity between LoRAs during analysis to account for directional alignment, then computes the exact vector-sum energy using the formula ||sum(v_i)||^2 = sum(||v_i||^2) + 2 * sum(||v_i|| * ||v_j|| * cos(v_i, v_j)). All strengths are scaled so the total combined energy matches what the strongest single LoRA would contribute alone.

• Aligned LoRAs (cos~1) — stronger reduction (they reinforce each other, so combined energy is high)
• Orthogonal LoRAs (cos~0) — moderate reduction (independent contributions add in quadrature)
• Opposing LoRAs (cos~-1) — minimal reduction (they cancel out, so combined energy is low)

Scenario	Result
2 aligned LoRAs (cos~1) at strength 1.0	Each reduced to ~0.50
2 orthogonal LoRAs (cos~0) at strength 1.0	Each reduced to ~0.71
2 opposing LoRAs (cos~-1) at strength 1.0	~1.0 each (they cancel)
1 strong + 1 weak LoRA	Proportional reduction
Single LoRA	No change
auto_strength disabled	No adjustment (default)

Your original strength ratios are always preserved — the algorithm only scales them down uniformly.

Architecture-Aware Key Normalization

Different LoRA trainers (Kohya, AI-Toolkit, LyCORIS, diffusers/PEFT) produce LoRAs with different key naming conventions for the same model weights. When mixing LoRAs from different trainers, the optimizer sees no key overlap and cannot merge them correctly.

Key normalization auto-detects the model architecture from LoRA key patterns and remaps all keys to a canonical format, enabling correct overlap detection and conflict analysis across trainer formats.

Architecture	Detected From	Normalization
Z-Image (Lumina2)	diffusion_model.layers.N.attention, single_transformer_blocks	Prefix standardization, QKV split for per-component analysis, re-fuse after merge
FLUX	double_blocks/single_blocks, transformer.transformer_blocks	AI-Toolkit / Kohya / diffusers unified to canonical format
Wan 2.1/2.2	blocks.N with self_attn/cross_attn/ffn	LyCORIS / diffusers / Musubi Tuner unified, RS-LoRA alpha fix
SDXL	lora_te1_/lora_te2_, input_blocks/down_blocks	Text encoder + UNet key unification
LTX Video	adaln_single, transformer_blocks with attn1/attn2	Trainer format unification
Qwen-Image	transformer_blocks with img_mlp/txt_mlp/img_mod/txt_mod	Dual-stream key unification

Z-Image QKV handling: Z-Image LoRAs often fuse Q, K, V projections into a single attention.qkv weight. The normalizer splits these into separate to_q/to_k/to_v components for per-component conflict analysis, then re-fuses them back to the native format after merging.

Setting	Default	Effect
normalize_keys	disabled	disabled or enabled. Enable when mixing LoRAs from different trainers or for Z-Image QKV fusion.

SVD Patch Compression

After merging, full-rank diff patches consume ~128x more RAM than standard LoRA patches (64MB vs 0.5MB per key for a 4096x4096 weight). The optimizer re-compresses merged patches to low-rank via truncated SVD, dramatically reducing post-merge RAM.

Mode	What gets compressed	Quality	RAM savings
non_ties (default)	weighted_sum and weighted_average prefixes only	Lossless — sum of input ranks preserves all merge information	~32x on compressed prefixes
all	Everything including TIES	Lossy on TIES prefixes — nonlinear ops (trim, sign election) produce full-rank results that can't be perfectly captured	~32x on all prefixes
disabled	Nothing	No loss	No savings

The compression rank is automatically computed as the sum of all input LoRA ranks. For example, 3 rank-32 LoRAs produce a rank-96 compressed patch — enough to represent the full merge without quality loss on linear operations.

Tip: For video models (LTX, Wan, etc.) with high RAM usage, use weighted_sum_only + non_ties (or all). Every patch gets losslessly compressed with minimal RAM footprint.

Optimization Modes

Mode	Behavior
per_prefix (default)	Each weight group picks its own strategy based on local conflict data
global	Single strategy for all prefixes (original behavior)
weighted_sum_only	Forces simple weighted sum everywhere — no TIES, no averaging. Combined with patch compression, all patches are fully compressible with zero quality loss

Block Strategy Map

The analysis report includes a visual block-by-block map showing what strategy was used and why:

--- Block Strategy Map ---
  input_blocks.0   ====  sum  1 LoRA (6x)
  input_blocks.4   ----  avg  12% conflict (6x)
  middle_block.1   ####  TIES 42% conflict (6x)
  output_blocks.3  ----  avg  8% conflict (6x)
  output_blocks.8  ====  sum  1 LoRA (6x)
  Legend: ==== sum (single LoRA)  ---- avg (compatible)  #### TIES (conflict)

Memory Options

Option	Default	Effect
cache_patches	enabled	Cache merged patches in RAM for faster re-execution. Disable to free RAM after merge (recommended for video models)
compress_patches	non_ties	SVD re-compression of merged patches (see above)
svd_device	gpu	Device for SVD compression. GPU is ~10-50x faster than CPU. Use CPU if GPU memory is tight
free_vram_between_passes	disabled	Release GPU cache between analysis and merge passes. Lowers peak VRAM at negligible speed cost

Inputs / Outputs

Inputs: MODEL, CLIP (optional), LORA_STACK, output strength, clip strength multiplier, auto strength, optimization mode, cache patches, compress patches, SVD device, free VRAM between passes, normalize keys, sparsification, sparsification density.

Outputs: MODEL, CLIP, STRING (analysis report), LORA_DATA (for Save Merged LoRA / Merged LoRA to Hook)

Example Report

==================================================
LORA OPTIMIZER - ANALYSIS REPORT
==================================================

--- Per-LoRA Analysis ---
  style_lora.safetensors:
    Strength: 1.0
    Keys: 192
    Avg rank: 64
    L2 norm (mean): 0.0847
  detail_lora.safetensors:
    Strength: 0.8
    Keys: 192
    Avg rank: 32
    L2 norm (mean): 0.0423

--- Auto-Strength Adjustment ---
  style_lora.safetensors: 1.0 -> 0.6345
  detail_lora.safetensors: 0.8 -> 0.5076
  Scale factor: 0.6345
  Method: interference-aware energy normalization
    Avg pairwise cosine similarity: 0.312 (mostly aligned (reinforcing))
    Interference-aware energy: 0.1335 (orthogonal assumption: 0.1196)

--- Pairwise Analysis ---
  style_lora.safetensors vs detail_lora.safetensors:
    Overlapping positions: 89420
    Sign conflicts: 31297 (35.0%)
    Cosine similarity: 0.312

--- Collection Statistics ---
  Total LoRAs: 2
  Total unique keys: 196
  Avg sign conflict ratio: 35.0%
  Magnitude ratio (max/min L2): 2.00x

--- Auto-Selected Parameters ---
  Merge mode: ties
  Density: 0.42
  Sign method: frequency
  Sparsification: DARE
  Sparsification density: 0.70 (keep rate)
  For TIES prefixes: replaces trim step; others: preprocessing
  (global fallback — each prefix uses its own parameters)

--- Per-Prefix Strategy ---
  weighted_sum (single LoRA):        28 prefixes (14%)
  weighted_average (low conflict):  120 prefixes (61%)
  ties (high conflict):              48 prefixes (24%)
  Total:                            196 prefixes

--- Block Strategy Map ---
  input_blocks.0   ====  sum  1 LoRA (6x)
  input_blocks.1   ====  sum  1 LoRA (6x)
  input_blocks.4   ----  avg  12% conflict (6x)
  input_blocks.5   ####  TIES 38% conflict (6x)
  middle_block.1   ####  TIES 42% conflict (6x)
  output_blocks.3  ----  avg  15% conflict (6x)
  output_blocks.8  ====  sum  1 LoRA (6x)
  Legend: ==== sum (single LoRA)  ---- avg (compatible)  #### TIES (conflict)

--- Reasoning ---
  Sign conflict ratio 35.0% > 25% threshold -> TIES mode selected
    TIES resolves sign conflicts via trim + elect sign + disjoint merge
  Auto-density estimated at 0.42 from magnitude distribution
  Magnitude ratio 2.00x <= 2x -> 'frequency' sign method (equal voting)
    Similar-strength LoRAs get equal votes

--- Merge Summary ---
  Keys processed: 196
  Model patches: 168
  CLIP patches: 28
  Output strength: 1.0
  CLIP strength: 1.0

==================================================

Connect the STRING output to a Show Text node to see the report in ComfyUI.

Structural & Edit LoRAs: Do not put distillation LoRAs (LCM, Lightning, Turbo, Hyper), DPO LoRAs, or edit model LoRAs (Qwen edit, Klein edit, instruction-editing LoRAs) in the optimizer stack. These LoRAs modify the model's fundamental behavior — their weights are precisely calibrated and merging them with style LoRAs can break their training. Apply them via a standard Load LoRA node upstream, then feed only your style/character LoRAs into the optimizer. If you must include an edit LoRA in the stack, use weighted_sum_only mode and disable sparsification to avoid weight trimming.

Limitation: The optimizer only analyzes LoRAs in its own stack. It cannot see LoRA patches applied by upstream nodes (Load LoRA, etc.) — those stack additively on top of the optimizer's output. Fully baked merges (safetensors checkpoints) are indistinguishable from base weights and cannot be detected.

---

Save Merged LoRA

Saves the optimizer's merged result as a standalone .safetensors file that works with any standard LoRA loader.

Connect the LORA_DATA output from LoRA Optimizer to this node.

Option	Default	Effect
filename	merged_lora	Plain name saves to your ComfyUI loras folder. Absolute path (e.g. /path/to/my_lora) saves to that location
save_rank	0 (auto)	0 = use each layer's existing rank from the merge. Non-zero = force this rank for layers that need compression
bake_strength	enabled	When on, the saved LoRA reproduces your exact merge at strength 1.0. When off, strengths are not baked in

Outputs: STRING (file path)

---

Merged LoRA to Hook

Wraps the optimizer's merged patches as a conditioning hook (HOOKS) for per-conditioning LoRA application. Instead of applying the merged LoRA globally to the model, you can attach it to specific conditioning entries using ComfyUI's hook system.

Connect the LORA_DATA output from LoRA Optimizer to this node, then connect the HOOKS output to a Cond Set Props (or similar) node.

Inputs: LORA_DATA (required), HOOKS (optional — chain with existing hooks)

Outputs: HOOKS

When to use this

Use this node when you want the merged LoRA to apply only to specific conditioning rather than the entire model:

• Per-prompt LoRA: Apply different merged LoRAs to positive vs negative conditioning
• Scheduled application: Combine with hook keyframes to apply the LoRA only during certain sampling steps
• Regional conditioning: Use with area-based conditioning to apply the LoRA to specific image regions
• Preserving the base model: Keep the MODEL output clean (unpatched) while still using the merged LoRA through conditioning hooks

Workflow example:

Load Checkpoint → MODEL ──┬──→ LoRA Optimizer → LORA_DATA → Merged LoRA to Hook → HOOKS
                           │                                                          ↓
                           └──→ KSampler ←──── Conditioning ←──── Cond Set Props

The prev_hooks input allows chaining multiple hook sources together.

---

WanVideo LoRA Optimizer

Variant of the LoRA Optimizer for WanVideo models (via kijai's WanVideoWrapper). Accepts WANVIDEOMODEL instead of MODEL, skips CLIP, and applies merged patches in-memory.

All merging algorithms are inherited — TIES, DARE/DELLA, SVD compression, auto-strength, per-prefix adaptive merge, and Wan key normalization (LyCORIS, diffusers, Fun LoRA, finetrainer, RS-LoRA) all work identically.

Inputs: WANVIDEOMODEL, LORA_STACK, output strength, and all optimizer options (except CLIP-related ones). Defaults: normalize_keys=enabled, cache_patches=disabled.

Outputs: WANVIDEOMODEL (patched), STRING (analysis report), LORA_DATA (for Save Merged LoRA)

Workflow and usage

Basic workflow:

WanVideoModelLoader → WANVIDEOMODEL → WanVideo LoRA Optimizer → WANVIDEOMODEL → WanVideoSampler
                                               ↑
                        LoRA Stack ─────────────┘

Chaining with individual LoRAs: Individual (non-merged) LoRAs go through WanVideoLoraSelect → model loader as usual. Our optimizer applies merged LoRAs on top — both coexist in the model patcher.

WanVideoLoraSelect → WanVideoModelLoader → WANVIDEOMODEL → WanVideo LoRA Optimizer → Sampler
                                                                    ↑
                                             LoRA Stack ────────────┘

Key defaults differ from the standard optimizer:

• normalize_keys = enabled — WanVideo LoRAs come from many trainers, normalization is commonly needed
• cache_patches = disabled — video models are large, caching uses significant RAM

---

Installation

ComfyUI Manager

Search for "LoRA Optimizer" in ComfyUI Manager and install.

Manual

cd ComfyUI/custom_nodes/
git clone https://github.com/ethanfel/ComfyUI-LoRA-Optimizer.git

Restart ComfyUI. Nodes appear under the loaders category.

Compatibility

• Models: SD 1.5, SDXL, Flux, Z-Image (Lumina2), Wan 2.1/2.2, LTX Video, Qwen-Image, and other architectures supported by ComfyUI
• LoRA formats: Standard LoRA, LoCon, LyCORIS, diffusers/PEFT formats
• Trainers: Kohya, AI-Toolkit, LyCORIS, Musubi Tuner, diffusers — auto-normalized when normalize_keys is enabled
• Flux sliced weights: Handled correctly (linear1_qkv offsets)
• Z-Image fused QKV: Split for per-component analysis, re-fused after merge
• Stack formats: Native LoRA Stack dicts, plus standard tuples from Efficiency Nodes / Comfyroll

Credits

• Originally based on ComfyUI-ZImage-LoRA-Merger by DanrisiUA
• Per-prefix adaptive approach inspired by comfyUI-Realtime-Lora by shootthesound (per-block LoRA analysis)
• Thanks to Scruffy and Ramonguthrie for suggesting the per-block analysis approach
• TIES-Merging: Yadav et al., NeurIPS 2023
• DARE: Yu et al., ICML 2024 — Drop And REscale for language model merging
• DELLA: Deep et al., 2024 — magnitude-aware sparsification

License

MIT License - see LICENSE.