ComfyUI Extension: LoRA Optimizer

Authored by ethanfel

Created 5 months ago

Updated 6 days ago

135 stars

Run ComfyUI workflows without the setup

No installs, no CUDA version roulette, no GPU sitting idle on your bill. Bring a workflow and run it in the browser.

Auto-optimizer node for combining multiple LoRAs via diff-based merging with TIES conflict resolution

README

Stacking multiple LoRAs in ComfyUI often causes oversaturation, artifacts, or lost details. This node suite automatically figures out the best way to combine your LoRAs — analyzing where they conflict, resolving those conflicts per model layer, and adjusting strengths so the result looks clean. Just add your LoRAs to a stack, connect the optimizer, and generate.

The Problem

<details> <summary>Should I Merge This LoRA?</summary> <img src="assets/merge-use-cases.png" width="720" alt="Should I merge this LoRA? Decision guide"> </details>

Merge Strategy Guide

How LoRAs relate to each other, what the optimizer does about it, and when to change settings.

Quick Start

Add a LoRA Stack (Dynamic) node — pick your LoRAs, set strengths
Add a LoRA Optimizer node — connect MODEL (and optionally CLIP) from your checkpoint
Connect the optimizer's MODEL/CLIP outputs to your sampler — done

LoRA Stack (Dynamic) ──► LoRA Optimizer ──► KSampler
                              ▲
Load Checkpoint ──► MODEL ────┘

Everything is automatic. Connect analysis_report to a Show Text node to see what the optimizer did.

Want more control? Add a Settings node → connect to the optimizer's settings input. Want the best config found for you? Use the LoRA AutoTuner instead of the optimizer.

Installation

ComfyUI Manager

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

<details> <summary>Manual install</summary>

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

Restart ComfyUI. Nodes appear under the loaders category.

</details>

Nodes at a Glance

| Node | What It Does | |------|-------------| | LoRA Stack / (Dynamic) | Build your list of LoRAs — pick files, set strengths | | LoRA Optimizer | Analyze + merge your stack automatically. Just connect and go | | LoRA Optimizer (Inline Chain) | Drop-in filter after regular Load LoRA nodes — merges their patches in place, no restacking | | LoRA Inline Chain Options | Optional side node for the Inline Chain optimizer — set per-LoRA enable/strength/conflict/preserve options | | Settings Nodes | Optional fine-tuning: sparsification, compression, smoothing, etc. | | LoRA AutoTuner | Sweep 2000+ parameter combos, rank the best configs | | LoRA Merge Estimator | Predict the best config via k-NN over the community cache — skip the sweep | | Merge Selector | Try alternative ranked configs from AutoTuner results | | Conflict Editor | Inspect pairwise conflicts, set per-LoRA conflict modes + merge strategy by hand | | Compatibility Analyzer | Check which LoRAs work well together before merging | | Merge Formula | Control hierarchical merge order, e.g. (1+2) + 3 | | Metadata Reader | Read embedded prompt / description / merge info from any LoRA | | Extract from Model | Recover a LoRA baked into a finetuned model by diffing against the base | | Combination Generator | Auto-generate 2-/3-way combos for AutoTuner dataset collection | | Save Merged LoRA | Export the merge as a standalone .safetensors file | | Merged LoRA to Hook | Apply merged LoRA per-conditioning instead of globally | | LoRA Optimizer (Legacy) | All parameters on one node (superseded by Optimizer + Settings) |

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

Full parameter reference: Nodes wiki page · Configuration Guide · Workflows

<details> <summary>Examples</summary>

Z-Image (Lumina2) — 3 LoRAs merged

How It Works

<a href="assets/optimizer-pipeline.png"><img src="assets/optimizer-pipeline.svg" alt="Optimizer Pipeline" width="100%"></a> <details> <summary>Per-Group Adaptive Merge — Deep Dive</summary>

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 resolved target group:

| Condition | Strategy | |-----------|----------| | Only 1 LoRA touches this group | weighted_sum — full strength, no dilution | | 2+ LoRAs, low excess conflict + low subspace overlap | weighted_average — mostly independent updates | | 2+ LoRAs, high similarity + low excess conflict | consensus — aligned, low-interference merge | | 2+ LoRAs, excess conflict > 25% with overlapping subspaces | ties — resolve real conflicts with trim/elect/merge | | Magnitude ratio > 2x in the group | total sign method (stronger LoRA dominates) | | Magnitude ratio <= 2x in the group | frequency sign method (equal votes) |

</div>

This means non-overlapping regions keep 100% of their LoRA's effect, while genuinely conflicting regions get proper TIES resolution. When decision_smoothing > 0, those per-group metrics are softly pulled toward the block average so adjacent layers do not flip strategies due to noisy samples.

Two-Pass Streaming Architecture

The optimizer uses a two-pass streaming architecture for low memory usage:

Pass 1 (Analysis): Resolves trainer aliases to target weights, aggregates alias collisions per LoRA, samples conflict and magnitude statistics per target group, then discards the diffs. Only lightweight scalars are kept.
Pass 2 (Merge): Recomputes diffs per target group, looks up that group's conflict data, picks a strategy for it, and merges. Each group is freed after merging. Standard linear merges stay in exact low-rank form; nonlinear merges and optional compression still use dense/SVD paths.

Peak memory is still roughly "one target group at a time," but the exact peak depends on the largest layer, how many LoRAs hit it, and whether extra quality/compression steps are enabled. GPU-accelerated on both passes.

<details> <summary>What It Analyzes</summary>

Per-LoRA metrics (rank, key count, effective L2 norms)
Pairwise raw + magnitude-weighted conflict ratios per target group (sampled for efficiency)
Excess conflict over the cosine baseline, plus low-rank subspace overlap
Pairwise cosine similarity (directional alignment between LoRAs)
Magnitude / activation-importance distribution per target group
Key overlap between LoRAs

</details> </details>

Optimizer Features

<details> <summary>TIES Merging</summary>

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.

<a href="assets/ties-diagram.png"><img src="assets/ties-diagram.svg" alt="TIES Merging Pipeline" width="100%"></a> </details> <details> <summary>DARE / DELLA Sparsification</summary>

DARE and DELLA sparsify each LoRA's diff before merging, reducing parameter interference between LoRAs. The implementations here are practical LoRA-oriented variants inspired by those papers, not paper-faithful reproductions. 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) | | dare_dampening | 0.0 | DAREx rescale dampening (0–1, ICLR 2025). Only affects dare / dare_conflict modes. 0 = standard DARE 1/density rescale. Higher values dampen the rescale to reduce noise amplification at low density |

</details> <details> <summary>Merge Refinement (Refine / Full)</summary>

Optional preprocessing steps applied to weight diffs before merging, selectable via the merge_refinement dropdown:

| Level | What It Adds | Cost | |-------|-------------|------| | none (default) | Merge as-is, no extra processing | Baseline | | refine | Direction orthogonalization + TALL-mask selfish weight protection | Minimal extra compute, no extra VRAM | | full | KnOTS SVD alignment + orthogonalization + TALL-masks | More VRAM for SVD decomposition |

TALL-masks (refine+): Identifies "selfish" weights — positions where one LoRA dominates and others contribute little. These weights are separated from the consensus merge and added back afterward, protecting each LoRA's unique features from being averaged away.

Direction orthogonalization (refine+): Projects LoRA diffs to be mutually orthogonal, reducing interference between LoRAs that modify overlapping weight regions.

KnOTS SVD alignment (full): Projects all LoRA diffs into a shared singular value basis via truncated SVD before merging. This makes diffs more directly comparable by aligning their representation spaces. Falls back to CPU on GPU OOM, skips gracefully if both fail.

Interaction with other settings:

Works with all merge modes (TIES, weighted_average, SLERP, etc.)
Combines with DARE/DELLA sparsification — sparsification runs first, then refinement
Best combination: maximum + della_conflict (or dare_conflict) for full pipeline
Single-LoRA prefixes: all enhancements short-circuit (no work to do)

| Setting | Default | Options | |---------|---------|---------| | merge_refinement | none | none, refine, full |

</details> <details> <summary>Key Filter</summary>

Each LoRA has a per-LoRA key_filter setting (available on both LoRA Stack and LoRA Stack (Dynamic) in advanced mode) that controls which target groups that LoRA contributes to, based on how many LoRAs in the stack share each resolved target:

| Filter | Behavior | Use Case | |--------|----------|----------| | all (default) | Contribute to all keys | Normal merging | | shared_only | Only contribute to keys present in 2+ LoRAs | Strip variant-specific keys (I2V/VACE) from this LoRA | | unique_only | Only contribute to keys present in exactly 1 LoRA | Extract only the variant-specific adapter keys from this LoRA | | audio_only | Only contribute audio layers | Take the sound from one LoRA on audio-video models (LTX-2, ACE-Step) | | no_audio | Only contribute non-audio (video) layers | Merge two LTX-2 LoRAs but keep just one's audio — set the others to no_audio |

Audio split (LTX-2 / ACE-Step): audio_only / no_audio classify a layer as "audio" when audio appears in its key (covers audio_embeddings_connector, audio_adaln_single, audio_patchify_proj, audio_proj_out, av_ca_audio_*, and the per-block audio sublayers). So to merge two action LoRAs but keep only the first one's sound, set the second LoRA's key_filter to no_audio. To combine an audio LoRA with a video LoRA, set the audio one to audio_only and the video one to no_audio (or all).

This is especially useful for Wan T2V/I2V/VACE LoRAs, which share ~90% of weights but each variant has unique keys (I2V: cross_attn.k_img/v_img, img_emb; VACE: vace_blocks.*, vace_patch_embedding).

Because the filter is per-LoRA, you can apply different filters to different LoRAs in the same stack — e.g., "take only the unique VACE keys from LoRA #2 while merging all keys from LoRA #1".

Example — making an I2V LoRA T2V-compatible:

Stack a T2V LoRA + an I2V LoRA together
Set the I2V LoRA's key_filter to shared_only
The I2V-only keys (k_img, v_img, img_emb, etc.) are skipped for that LoRA since they appear in only 1 LoRA
The merged result contains only the shared T2V-compatible weights

Example — extracting a lightweight I2V adapter:

Same stack (T2V + I2V)
Set the I2V LoRA's key_filter to unique_only
Only the I2V-specific keys are contributed by that LoRA — a small adapter with just the variant-specific weights

The filter uses the raw n_loras count from Pass 1 (before any filtering) and now participates in analysis as well as Pass 2 merge.

</details> <details> <summary>Auto-Strength</summary>

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 strong merge settings.

The algorithm uses interference-aware energy normalization: during Pass 1 it streams exact Frobenius norms and pairwise dots for each LoRA branch, then computes the exact vector-sum energy separately for model and CLIP updates. All strengths are uniformly 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, optionally clamped by an architecture-aware floor
Opposing LoRAs (cos~-1) — minimal reduction (they cancel out, so combined energy is low)

When orthogonal LoRAs are effectively independent, the optimizer can clamp the scale factor with auto_strength_floor:

| Architecture | Default floor | |-------------|---------------| | Wan / LTX Video | 1.0 | | SD / SDXL / Flux / Z-Image | 0.85 | | LLM-style presets | 0.9 |

auto_strength_floor = -1 uses the architecture default. Setting 0.0–1.0 overrides it manually.

| 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 before floor-clamping | | 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 |

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

</details> <details> <summary>Output & CLIP Strength</summary>

The LoRA Optimizer has two strength inputs that control the merged result globally:

| Input | Default | Effect | |-------|---------|--------| | output_strength | 1.0 | Master volume for the merged result. 1.0 = full effect, 0.5 = half, 0 = disabled. Set to -1 for auto — the optimizer picks the suggested max strength, compensating for energy lost during the merge | | clip_strength_multiplier | 1.0 | How strongly the LoRAs affect text understanding (CLIP). At 1.0 it matches the model strength; lower values reduce LoRA influence on prompts while keeping the visual effect. Leave at 1.0 (or ignore) for video models with no CLIP |

output_strength = -1 is the easiest way to let the optimizer fully drive strength: combined with auto_strength, it both balances per-LoRA contributions and sets a sensible master level for the stack.

</details> <details> <summary>Decision Smoothing</summary>

decision_smoothing — blends each group's decision metrics toward the average of its surrounding block. This reduces jagged layer-to-layer mode flips when the stack is noisy.

smooth_slerp_gate — when enabled, uses per-prefix cosine similarity (computed during analysis) instead of the collection average for the SLERP interpolation gate. This makes the SLERP weight vary per layer based on local alignment rather than using a single global value. Available on the LoRA Merge Settings node or the Legacy optimizer.

</details> <details> <summary>Architecture-Aware Key Normalization</summary>

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 | | Ideogram 4 | layers.N.attention.qkv/attention.o, feed_forward.w1-w3, fal conditional_transformer. prefix | ai-toolkit / fal / PEFT prefixes unified; qkv stays fused (native ComfyUI layout) | | 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 | | ACE-Step | layers.N with self_attn/cross_attn and q_proj/k_proj/v_proj | Attention key unification | | Anima (Cosmos-Predict2 DiT) | blocks.N.{self_attn,cross_attn}.{q,k,v,output}_proj, mlp.layer1/2, unique llm_adapter; Kohya lora_unet_* / diffusers transformer_blocks.attn1/attn2 | Kohya / diffusers / ComfyUI unified to diffusion_model.blocks.N.*; split QKV | | 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 | enabled | disabled or enabled. Recommended for mixed-trainer stacks and required for Z-Image QKV fusion. |

</details> <details> <summary>Architecture-Aware Behavior Profiles</summary>

All numeric thresholds in the optimizer (density estimation, conflict detection, auto-strength scaling, scoring heuristics) are tuned per architecture family. The architecture_preset setting selects the appropriate thresholds — auto detects from LoRA key patterns.

| Preset | Architectures | Key Differences | Orthogonal floor | |--------|--------------|-----------------|------------------| | sd_unet | SD 1.5, SDXL | Density range [0.1, 0.9], noise floor 10%, max strength cap 3.0 | 0.85 | | dit | Flux, WAN, Z-Image, LTX, Ideogram 4, Anima, HunyuanVideo | Density range [0.4, 0.95], noise floor 5%, max strength cap 5.0 | 0.85 by default, 1.0 for Wan/LTX | | acestep_dit | ACE-Step (music DiT) | DiT thresholds tuned for music LoRAs: wider orthogonal band + higher TIES threshold to preserve voice/timbre | 1.0 | | llm | Qwen-Image, LLaMA-based | Density range [0.1, 0.8], noise floor 15%, max strength cap 3.0 | 0.9 |

Why it matters: DiT architectures have denser weight distributions than UNet — with UNet thresholds, the optimizer underestimates density and clips suggested strength too aggressively. LLM-based models are sparser and benefit from lower density ceilings. ACE-Step gets its own preset because music LoRAs are unusually conflict-prone and need a gentler merge to keep the singing voice intact.

| Setting | Default | Options | |---------|---------|---------| | architecture_preset | auto | auto, sd_unet, dit, acestep_dit, llm. Auto-detection uses the same key pattern matching as key normalization |

HunyuanVideo: covered by the dit preset's thresholds, but it is not in the auto-detector — select dit manually for HunyuanVideo LoRAs.

Note: This is orthogonal to strategy_set (which controls which strategies are available — consensus, SLERP, etc.). Architecture preset controls the numeric thresholds those strategies use.

</details> <details> <summary>SVD Patch Compression</summary>

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 | |------|---------------------|---------|-------------| | smart (default) | weighted_sum and weighted_average prefixes only | Lossless — sum of input ranks preserves all merge information | ~32x on compressed prefixes | | aggressive | 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 |

When dense compression is needed, 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 on linear operations when no extra nonlinear processing is involved.

Tip: For video models (LTX, Wan, etc.) with high RAM usage, use additive mode + smart (or aggressive) compression. Every patch gets losslessly compressed with minimal RAM footprint.

</details> <details> <summary>Optimization Modes</summary>

| 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) | | additive | Simple weighted addition — no conflict resolution. Preserves all weights exactly. Use for DPO/edit/distill LoRAs, or with patch compression for minimal RAM |

</details> <details> <summary>Block Strategy Map</summary>

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)

</details> <details> <summary>Memory Options</summary>

| 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) | | patch_compression | smart | 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 |

</details> <details> <summary>Example Report</summary>

==================================================
LORA OPTIMIZER - ANALYSIS REPORT
==================================================
Architecture preset: sd_unet (SD/SDXL UNet)

--- 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.

</details> <details> <summary>Important notes & limitations</summary>

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 additive 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. To capture and merge an existing Load LoRA chain instead, use the LoRA Optimizer (Inline Chain) node. Fully baked merges (safetensors checkpoints) are indistinguishable from base weights and cannot be detected.

</details>

AutoTuner

Automatically sweeps all merge parameters (mode, sparsification, density, dampening, quality level) and ranks configurations for your LoRA stack. Runs Pass 1 analysis once, scores all parameter combinations via heuristic proxies, then merges the top-N candidates and measures output quality. When an AUTOTUNER_EVALUATOR is connected, the built-in score can be blended with external prompt/reference evaluation logic. Outputs the highest-ranked merge directly as MODEL/CLIP, plus a ranked report and TUNER_DATA for exploring alternatives via a Merge Selector node.

Full parameter reference: Nodes wiki page

<details> <summary>Diff Cache</summary>

During the parameter sweep, each candidate recomputes raw LoRA diffs (A@B matmul) from scratch — even though diffs depend only on LoRA content, not merge config. The diff cache stores these diffs after the first candidate and reuses them for subsequent candidates, eliminating redundant computation.

| Mode | Behavior | |------|----------| | disabled | Recomputes diffs each time. No extra memory | | auto | Uses RAM up to diff_cache_ram_pct of free memory, then spills to disk. Recommended for most setups | | ram | All diffs in RAM. Fastest, but uses ~1.5 GB (SDXL) to ~6 GB (Flux) | | disk | All diffs to temp files with memory-mapping. Slowest cache mode, but minimal RAM |

Disk bound: the disk spill (auto/disk) is capped by free space — it keeps ~5 GB free on the temp volume, then stops caching and recomputes the overflow diffs on demand instead of filling the disk. Large models with several LoRAs (e.g. LTX-2 audio-video × 3) can otherwise spill tens of GB of full-rank diffs to ComfyUI/temp/. If you hit the cap a lot, point ComfyUI's temp dir at a larger drive or use diff_cache_mode=disabled.

| Setting | Default | Effect | |---------|---------|--------| | diff_cache_mode | auto | Cache mode selection | | diff_cache_ram_pct | 0.5 | Fraction of free system RAM for auto mode (0.1–0.9) |

</details> <details> <summary>Scoring Modes & Speed</summary>

The AutoTuner ranks candidate configs with a heuristic score before merging the top-N for real quality measurement. These knobs trade scoring accuracy for speed:

| Setting | Default | Options / Effect | |---------|---------|------------------| | scoring_speed | turbo | How many prefixes each candidate is scored on. full = all prefixes (slowest, most accurate); fast ≈ every 2nd; turbo ≈ every 3rd (recommended); turbo+ = fastest, biased toward high-conflict prefixes | | scoring_svd | disabled | SVD-based scoring. disabled = fast norm-only (usually enough); merge_quality = SVD on merged diffs (more thorough); lora_rank = effective-rank of factors (experimental, changes ranking); full = both. Hardware-accelerated when Triton is installed | | scoring_device | gpu | Where scoring math runs. gpu is much faster, especially with SVD modes | | scoring_formula | v2 | v2 (recommended) = arch-aware sparsity + energy metrics. v1 = legacy formula with a fixed 40% sparsity target, kept for comparison |

Ranking stays fair because every candidate is scored on the same subset of prefixes.

</details> <details> <summary>Persistent Memory</summary>

memory_mode caches a stack's tuning result across ComfyUI sessions, keyed by LoRA content hash + config. Re-running the same stack then skips the whole sweep and replays the winning config instantly.

| Mode | Behavior | |------|----------| | disabled | Never read or write the memory store | | auto (default) | Load cached result if present, save after tuning | | auto_ignore_strength | Same as auto, but the cache key ignores LoRA strengths — useful when sweeping strengths on orthogonal LoRAs where the ranking doesn't change | | read_only | Use cached results but never save new ones | | clear_and_run | Delete the cached entry and re-tune from scratch |

record_dataset (disabled / enabled) — when enabled, appends analysis metrics and all scored configs to user/lora_optimizer_reports/autotuner_dataset.jsonl for threshold-tuning research. Entries are written only when a full sweep actually runs (cache/memory replays add nothing).

</details> <details> <summary>VRAM Budget</summary>

The vram_budget slider (0.0–1.0) controls what fraction of free VRAM to use for storing merged patches on GPU. Default is 0 (all patches on CPU). Setting it higher keeps patches on GPU, reducing RAM usage on systems with enough VRAM. Available on both LoRA Optimizer and LoRA AutoTuner.

</details> <details> <summary>Community Cache</summary>

LoRA analysis results (conflict metrics, per-LoRA stats, best merge configs) are hardware-agnostic — the same LoRA files always produce the same output regardless of GPU. The community cache lets any user download precomputed results for their LoRAs without running the AutoTuner sweep, and optionally contribute their own results back.

Results are keyed by content hash (SHA256 of file contents, not filename), so they match across different users and folder layouts. LoRA names and paths are never shared.

| community_cache value | Behavior | |---|---| | disabled | No community interaction (default) | | upload_only | Runs the sweep locally and uploads results, but does not replay HF cache hits. Useful for backfilling/enriching configs | | upload_and_download | Downloads before analysis; uploads after if local score is higher |

Downloads are anonymous (no setup required). Uploads require a HF_TOKEN environment variable with write access to the dataset repo.

Results are stored in the public dataset ethanfel/lora-optimizer-community-cache.

</details>

LoRA Merge Estimator

Skip the AutoTuner sweep when your stack is similar to combos already in the community cache. The Estimator analyzes your LoRAs once (Phase 1 only), then retrieves the k nearest combos from a prebuilt index of cached configs and emits TUNER_DATA with the aggregated top-N predicted configs. Feed that directly into the LoRA Optimizer (settings_source=from_tuner_data) to apply the predicted merge.

Workflow:

LoRA Stack (Dynamic) ──► LoRA Merge Estimator ──► TUNER_DATA ──► LoRA Optimizer
                              ▲                                   (settings_source
Load Checkpoint ──► MODEL ────┘                                    =from_tuner_data)

The first run downloads the community cache and builds a local k-NN index under ComfyUI/models/estimator/ (~30–60s). Subsequent runs reuse the index and complete in seconds — no Phase 2 sweep, no merge quality pass.

| Input | Default | Effect | |------|---------|--------| | model / lora_stack | — | Same as optimizer/AutoTuner | | clip (optional) | — | Only used by Phase 1 analysis | | k | 5 | How many nearest neighbors to retrieve (1–20) | | rebuild_index | auto | auto rebuilds when the HF dataset or index schema changes; force always rebuilds; skip never rebuilds (errors if missing) | | top_n_output | 3 | How many aggregated candidates to emit in TUNER_DATA |

Outputs TUNER_DATA (consumable by the optimizer or Merge Selector) and a human-readable estimator_report listing predicted configs with their aggregated scores and retrieved neighbor distances.

When to use Estimator vs AutoTuner:

Estimator — fast prediction (seconds after the first run). Best when your stack resembles combos others have already tuned, or when you want a quick starting point before reaching for the full sweep.
AutoTuner — authoritative grid search with measured quality scoring. Best for novel combos, uncommon architectures, or when accuracy matters more than speed.

The two are complementary: try the Estimator first, and fall back to AutoTuner if the predicted configs underperform or no close neighbors exist (the report will say No neighbors when the index has no matching family+combo-size rows).

<details> <summary>Other Nodes & Workflows</summary>

Merge Selector

Applies a specific configuration from AutoTuner results without re-running the sweep. Connect TUNER_DATA from a LoRA AutoTuner (or Load Tuner Data) node and set the selection index to choose which ranked configuration to apply (1 = top-ranked, 2 = next-ranked, etc.).

Workflow:

LoRA AutoTuner → TUNER_DATA → Merge Selector (selection=2) → try the 2nd-ranked config
                      ↓
              Save Tuner Data → (reload later) → Load Tuner Data → Merge Selector

Conflict Editor

Inserts between a LoRA Stack and the LoRA Optimizer to give you manual control over conflict handling. It loads every LoRA, computes full-rank diffs, measures pairwise sign disagreement, and auto-suggests a per-LoRA conflict_mode plus an overall merge strategy. You can accept the suggestions or override each LoRA's mode by hand.

Inputs: lora_stack, merge_strategy (auto, ties, consensus, slerp, weighted_average, weighted_sum), and a per-slot conflict_mode_1..10 (auto, all, low_conflict, high_conflict)
Outputs: the enriched LORA_STACK, a human-readable analysis_report, and the resolved merge_strategy string

Use it when you already know two LoRAs fight and want to force high_conflict (TIES) handling on one of them, or to read the conflict report before committing to a merge.

Merge Formula

Passthrough node that attaches a hierarchical merge order to the stack. The formula uses numbered LoRA positions, + for blends, and parentheses for grouping — e.g. (1+2) + 3 merges LoRAs 1 and 2 first, then blends the result with LoRA 3. Optional per-component weights are supported: (1+2):0.6 + 3:0.4. An empty formula falls back to a flat merge of the whole stack.

Inputs: lora_stack, formula (string)
Outputs: lora_stack (unchanged tensors, formula metadata attached)

Useful when a clean two-LoRA base should be established before a third, more disruptive LoRA is layered on top.

Metadata Reader

Passthrough node that extracts embedded metadata from every LoRA in the stack — works with any .safetensors LoRA, not just ones saved by this pack. It consolidates fields like source_loras, merge_mode, merge settings, training prompt, and description.

Inputs: lora_stack
Outputs: lora_stack (unchanged), prompt, description, metadata_info

Connect the string outputs to Show Text nodes to inspect what a downloaded or previously merged LoRA actually contains.

Extract from Model

Recovers a LoRA that was baked into a finetuned model by subtracting the original base-model weights and SVD-decomposing the per-layer delta. Can extract both UNet and CLIP layers when the matching base/finetuned CLIP pair is provided.

| Input | Default | Effect | |-------|---------|--------| | base_model / finetuned_model | — | The clean base and the finetuned checkpoint to diff | | base_clip / finetuned_clip (optional) | — | Provide both to also extract CLIP-side layers | | rank | 32 | Target rank when rank_mode = fixed | | rank_mode | auto | auto picks rank per layer from energy_threshold; fixed forces rank | | energy_threshold | 0.99 | Fraction of singular-value energy to retain in auto mode | | strength | 1.0 | Strength baked into the emitted stack entry |

Outputs: LORA_STACK (feed straight into the Optimizer/AutoTuner) and LORA_DATA (feed into Save Merged LoRA to write a .safetensors)

Combination Generator

Generates LoRA combinations for AutoTuner dataset collection. It cycles through all 2-way and/or 3-way combos with deterministic shuffling and tracks progress in a persistent combo_progress.json, so a long collection run can be resumed without repeating combos.

| Input | Default | Effect | |-------|---------|--------| | shuffle_order | 0 | Seed for deterministic shuffle ordering | | strength | 1.0 | Strength applied to each LoRA in the emitted combo | | combo_size | 2_and_3 | 2, 3, or 2_and_3 | | folder_filter | "" | Comma-separated path prefixes to restrict the pool (e.g. zit/,zib/) | | rerun_mode | false | Re-emit already-processed combos for backfilling enriched configs | | rerun_source | shuffle | shuffle or original_progress ordering when re-running |

Outputs: one LORA_STACK at a time plus a combo_info string. Drive it with AutoTuner record_dataset / community_cache to build the shared cache.

AutoTuner → Optimizer Bridge

Chain the AutoTuner and the Legacy optimizer in a single model line for a "rank, then tweak" workflow. Only one node merges at a time — the other passes the model through. The Legacy optimizer's settings_source switch controls which node is authoritative, and the UI bridge keeps the paired widgets in sync.

[Load Model] → [AutoTuner] → model → [Optimizer (Legacy)] → MODEL → sampler
[LoRA Stack]  → [AutoTuner]
[LoRA Stack]  → [Optimizer (Legacy)]
               [AutoTuner] → tuner_data → [Optimizer (Legacy)]

| Legacy Optimizer settings_source | What happens | |----|----| | from_autotuner | AutoTuner merges → Legacy Optimizer passes through. Optimizer widgets show the winning config. | | manual | AutoTuner passes the base model through → Legacy Optimizer merges with its own widget settings. | | from_tuner_data | Legacy Optimizer reads settings from connected tuner_data input. |

Typical flow:

Start with from_autotuner — let the AutoTuner find the best config
Inspect the Optimizer's widgets to see what won
Switch to manual — the Optimizer takes over, starting from the AutoTuner's recommendation
Tweak settings (merge_refinement, sparsification, etc.) and re-run

Switching between modes is instant — the AutoTuner reuses its cached sweep results.

Note: The bridge workflow requires the LoRA Optimizer (Legacy) node. The simplified LoRA Optimizer uses Settings nodes and tuner_data input instead.

Save / Load Tuner Data

Two utility nodes for persisting AutoTuner results to disk:

Save Tuner Data — Saves TUNER_DATA into a selected tuner_data folder as .tuner or .json. Subdirectories are allowed; path traversal outside that folder is blocked. Optional overwrite control avoids clobbering previous runs. OUTPUT_NODE = True.

Load Tuner Data — Dropdown of saved tuner data files. Outputs TUNER_DATA ready for Merge Selector. Auto-reloads when the file changes on disk.

Evaluator Utilities

Build AutoTuner Python Evaluator — packages a Python module path + callable name into an AUTOTUNER_EVALUATOR object. The callable can run prompts, compare references, and return a score in [0, 1].

The evaluator callable receives keyword arguments: model, clip, lora_data, config, context, and analysis_summary.

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 | |--------|---------|--------| | save_folder | first configured LoRA folder | Choose which configured ComfyUI LoRA directory to save into | | filename | merged_lora | File name relative to save_folder. Subdirectories are allowed (e.g. merged/my_lora) | | 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

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.

LoRA Optimizer (Inline Chain)

Drop-in filter for workflows built on regular Load LoRA nodes — no restacking required. Place it after your loader chain: it reads the LoRA patches the loaders left on MODEL/CLIP, strips the originals, merges them with the same optimizer engine, and re-applies the merged result. Outputs match the other optimizer nodes (model, clip, analysis_report, tuner_data, lora_data), so Save Merged LoRA chaining works.

Load Checkpoint ──► Load LoRA #1 ──► Load LoRA #2 ──► ... ──► LoRA Optimizer (Inline Chain) ──► KSampler
                                                                      ▲   (reads, strips, merges, re-applies)
                                          LoRA Inline Chain Options ──┘ chain_options (optional)

When to use it: you already have a workflow full of Load LoRA nodes and want the optimizer's conflict-resolved merge without rebuilding it around a LoRA Stack. For new workflows, the Stack → Optimizer path is still preferred — loading from files gives the optimizer real names, metadata, and architecture auto-detection.

Per-LoRA options live on the side node. Connect a LoRA Inline Chain Options node to the inline node's chain_options input to set per-LoRA enable/strength/conflict/preserve options. Leave chain_options unconnected to merge every captured LoRA with default options — the inline node works standalone.

Slot order = chain order. Slot #1 (on the options node) is the first Load LoRA in the chain (closest to the checkpoint). The report opens with per-LoRA chain fingerprints so you can verify the attribution:

[Inline Optimizer] Detected loader chain (slot -> LoRA mapping):
  #1: my_style.safetensors — 210 keys, rank 16, loader strength 0.80
  #2: my_char.safetensors — 176 keys, rank 32, loader strength 0.50, +48 clip keys @ 0.50

When the chain is fed by stock Load LoRA / Load LoRA (Model Only) nodes or the rgthree Power Lora Loader, the inline node recovers the real LoRA filenames and shows them in the fingerprint (as above). Other loaders that don't tag their output fall back to a generic chain lora #N label.

| Per-slot option | Effect | |---|---| | enabled | Off = that LoRA is removed from the model entirely (not merged, not applied) | | strength (simple) | Multiplier on the loader's strength — 1.0 keeps what the loader set; it is not an override | | model_strength / clip_strength (advanced) | Separate multipliers for the model and text-encoder branches | | conflict_mode / key_filter / preserve (advanced) | Same per-LoRA controls as LoRA Stack (Dynamic) |

Non-LoRA patches on the incoming model — OFT/BOFT rotations, hooked entries, padded diffs, third-party patch shapes — pass through untouched and are counted in the report. A Settings node is supported via the settings input in both advanced mode and AutoTuner mode — the AutoTuner's multi-candidate search, memory, and community cache all run on captured chains.

What that means for memory and community caching inline:

Stock and rgthree loaders reconcile with the file-based dataset. When a chain is fed by stock Load LoRA / Load LoRA (Model Only) or the rgthree Power Lora Loader, the inline node recovers each real filename and keys memory + community on the file bytes — the exact same identity a LoRA Stack run of that file uses. So a stack you tune inline shares its memory and community-cache entries with the Stack path, and vice-versa. Per-LoRA attribution is exact.
Only genuinely unstamped loaders use the separate captured namespace. A custom loader that doesn't tag its output leaves no filename, so those LoRAs fall back to a content hash of the captured weights (not the file bytes). Such chains share configs with other inline chains of the same captured content, but not with file-based recordings — this is the fallback, not the default for common loaders.
Memory persists across sessions either way. Both the file identity (stamped loaders) and the captured-content identity (unstamped) are stable across ComfyUI restarts, so a chain you tuned once is found again next run — unlike the per-session capture names, which change every restart.
The captured-namespace fallback assumes the same ComfyUI version. Its factors and key names depend on comfy's key-mapping / QKV-fusion, so a captured-content hash is only comparable across machines running the same comfy build. (File-identity reconciliation via stamped names is not affected — it hashes the file bytes.)

v1 limitations:

Fully-disjoint LoRAs at identical strengths may be grouped or ordered ambiguously — the chain fingerprints reveal it when it happens.
Model↔CLIP pairing is by chain order — a LoraLoaderModelOnly or a TE-only LoRA file in the chain can shift clip attribution; the report warns when the model/clip group counts differ.
Architecture detection reports unknown for inline capture — set architecture_preset via a Settings node to get arch-tuned thresholds.
WanVideo wrapper models are not supported — use the dedicated WanVideo LoRA Optimizer below.

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, merge refinement (KnOTS, orthogonalization, TALL-masks), and Wan key normalization (LyCORIS, diffusers, Fun LoRA, finetrainer, RS-LoRA) all work identically.

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
architecture_preset = dit — DiT-tuned thresholds (higher density floor, wider strength range)

</details>

<details> <summary>Compatibility</summary>

Models: SD 1.5, SDXL, Flux, Z-Image (Lumina2), Ideogram 4, Anima (Cosmos-Predict2), Wan 2.1/2.2, LTX Video, ACE-Step, Qwen-Image, and other architectures supported by ComfyUI
LoRA formats: Standard LoRA, LoCon, and LoRA/LoCon-style trainer variants whose tensors reduce to up/down(/mid) adapters (including many diffusers/PEFT and LyCORIS naming schemes)
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

</details> <details> <summary>Credits</summary>

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
KnOTS: Ramé et al., 2024 — SVD alignment for model merging
TALL-masks: Wang et al., 2024 — selfish weight protection via task-aware masks
Column-wise merging inspired by ZipLoRA: Shah et al., 2025 — structural sparsity for LoRA merging
Per-LoRA cleaning (STAR spectral truncate + rescale, base-norm magnitude taming) inspired by comfyui-lora-loader by CoreyCorza — per-LoRA noise-tail removal + layer taming as a preprocessing pass
STAR: Spectral Truncation And Rescale, NAACL 2025 — nuclear-norm-preserving spectral truncation for model merging
Norm-Anchor Scaling: arXiv:2602.02543 — magnitude taming anchored to the base-model reference norm

</details> <details> <summary>Development Timeline</summary> <a href="assets/timeline.png"><img src="assets/timeline.svg" alt="Development Timeline" width="720"></a> </details>

License

GPL-3.0 License - see LICENSE.

Run ComfyUI workflows without the setup

No installs, no CUDA version roulette, no GPU sitting idle on your bill. Bring a workflow and run it in the browser.

Learn more

ComfyUI Extension: LoRA Optimizer

Custom Nodes (25)

README

The Problem

Merge Strategy Guide

Quick Start

Installation

ComfyUI Manager

Nodes at a Glance

Z-Image (Lumina2) — 3 LoRAs merged

How It Works

Two-Pass Streaming Architecture

Optimizer Features

AutoTuner

LoRA Merge Estimator

Merge Selector

Conflict Editor

Merge Formula

Metadata Reader

Extract from Model

Combination Generator

AutoTuner → Optimizer Bridge

Save / Load Tuner Data

Evaluator Utilities

Save Merged LoRA

Merged LoRA to Hook

LoRA Optimizer (Inline Chain)

WanVideo LoRA Optimizer

License

Run ComfyUI workflows without the setup