Random LoRA Loader for ComfyUI

A ComfyUI custom node package for randomly selecting and applying LoRAs. Includes three nodes:

Random LoRA Loader - Select from 3 folders simultaneously
Filtered Random LoRA Loader - Single folder with keyword filtering
Filtered Random LoRA Loader (LBW) - 🆕 LBW support with automatic SD1.5/SDXL detection (NEW v1.2.0)

⚠️ Important: Supported Models

These nodes are designed for SD1.5 and SDXL models ONLY.

✅ Supported: Stable Diffusion 1.5, Stable Diffusion XL (SDXL)
❌ NOT Supported: Flux, SD3, SDXL Turbo, Pony, or other architectures

The LBW node's block weight feature specifically targets SD1.5/SDXL U-Net architecture. Other model types will not work correctly with LBW.

Nodes Overview

Random LoRA Loader (Original)

Select LoRAs from up to 3 different folders. Ideal for folder-based organization.

Use case:

Different folders for styles, characters, concepts
Fixed LoRA categories
Simple folder-based workflow

Filtered Random LoRA Loader

Select LoRAs from a single folder using keyword filtering. Ideal for dynamic selection and large collections.

Use case:

All LoRAs in one folder, filter by keywords
Dynamic filtering with AND/OR conditions
Metadata search capability
Recommended for chaining multiple instances

Filtered Random LoRA Loader (LBW) 🆕 NEW v1.2.0

Advanced node with LoRA Block Weight (LBW) support for precise effect control.

Use case:

Separate control of style vs. structure
Fine-tuned LoRA application
Professional workflows requiring precision
4 preset modes + custom input

See README_LBW.md for detailed LBW documentation.

Key Features

Common Features (All Nodes)

Multi-Source Metadata Reading: Automatically retrieve trigger words and sample prompts with priority order:
1. .metadata.json format (ComfyUI Lora Manager) - Priority 1
2. .info format (Civitai Helper) - Priority 2
3. Embedded metadata in LoRA file - Priority 3
Strength Randomization:
- Range specification with negative support: 0.4-0.8, -0.8--0.3 (v1.2.0)
- 0.1 increments for ranges: Easier to observe effect differences
- Fixed values rounded to 2 decimals: 0.847 → 0.85
Flexible Trigger Word Retrieval:
- json_combined: Combine all trigger word patterns (with deduplication)
- json_random: Randomly select one pattern
- json_sample_prompt: Randomly retrieve sample prompts
- metadata: Read directly from embedded metadata
Sample Prompt Optimization: Automatically remove LoRA syntax from samples and apply node-configured strength
Text Outputs: positive_text / negative_text, plus lora_text carrying the LoRA syntax (v1.4.0)
ComfyUI Standard Seed Control: Supports fixed/randomize/increment/decrement
Wildcard Encode Integration: Works seamlessly with Wildcard Encode (Inspire)
LoRA Syntax Auto-Removal: Cleans up LoRA syntax in additional_prompt to prevent noise
Preview Image Support: Display preview images/videos for selected LoRAs (v1.1.0)
Duplicate Filename Handling: Exclude duplicate filenames across subfolders (v1.1.0)

Filtered Nodes (Filtered & LBW)

Keyword Filtering: Space-separated keywords with phrase support
AND/OR/OFF Modes: Flexible filtering logic (OFF: v1.5.0)
Metadata Search: Search in filenames or embedded metadata
Fast Caching: Instant metadata access after first load
Keyword Output: Outputs the keywords the selected LoRA actually matched (v1.5.0)

LBW Node Only 🆕

LoRA Block Weight Support: Control which U-Net blocks are affected
Automatic SD1.5/SDXL Detection: No manual model type selection needed
4 Preset Modes:
- Style Focused - OUTPUT blocks only
- Character Focused - Balanced IN+MID+OUT
- Structure/Composition Only - INPUT+MID blocks only
- Balanced / Soft - Gentle application
Preset: Random Mode: Randomly select one of the 4 presets
Direct Input Mode: Custom weight specification
Automatic Weight Adjustment: Handles mismatched element counts

Installation

⚠️ Optional: Video Preview Support

For video file preview (.mp4, .webm, .avi, .mov), opencv-python is REQUIRED:

pip install opencv-python

Without opencv-python:

✅ Static images (.png, .jpg, .jpeg) work
✅ Animated images (.gif, .webp) work
❌ Video files will show BLACK SCREEN (with warning message in console)

LoRA functionality works normally regardless of opencv-python installation. Only preview display is affected.

Requirements

ComfyUI (works with standard installation)
Python 3.9+
Core dependencies included with ComfyUI ✅

Steps

cd ComfyUI/custom_nodes
git clone https://github.com/YOUR_USERNAME/RandomLoRALoader.git
# Or manually create RandomLoRALoader folder and copy files

Restart ComfyUI.

Note: Uses only Python standard library and ComfyUI bundled libraries for core functionality. opencv-python is optional for video preview only.

Uninstallation

# Navigate to custom_nodes folder
cd ComfyUI/custom_nodes

# Remove RandomLoRALoader folder
rm -rf RandomLoRALoader

# On Windows
rmdir /s RandomLoRALoader

Or manually delete the RandomLoRALoader folder and restart ComfyUI.

Usage

Basic Usage (Random LoRA Loader)

Add Node: Search for "Random LoRA Loader" in the node browser
Connect MODEL/CLIP: Connect base model and CLIP to inputs
Set Folder Paths:
- Group 1: Style LoRA folder path
- Group 2: Character LoRA folder path (optional)
- Group 3: Concept LoRA folder path (optional)
Configure Each Group:
- num_loras: Number of LoRAs to select from each group
- model_strength / clip_strength: Application strength (fixed value or random range)
Connect Outputs:
- MODEL / CLIP: Connect to KSampler
- positive / negative: Connect to Set Conditioning or KSampler
- positive_text / negative_text: Connect to Show Text for verification (recommended)
- lora_text: Connect to Show Text or a prompt/metadata saving node to record which LoRAs were picked

Wildcard Encode Integration (Recommended)

This node works seamlessly with Wildcard Encode (Inspire) for dynamic prompt generation combined with random LoRA selection.

Recommended Connection

[Wildcard Encode (Inspire)]
  text: "__style__, {red|blue|green}, <lora:base_effect:0.5>"
  ↓
  ├─ populated_text ──→ [Random LoRA Loader]
  │                     additional_prompt_positive
  └─ MODEL/CLIP ──────→ model/clip input
  
[Random LoRA Loader]
  Folder 1: character LoRAs
  Folder 2: concept LoRAs
  ↓
  MODEL/CLIP/CONDITIONING → [KSampler]

How It Works

Wildcard Encode Processing:
- Wildcard expansion: __style__ → "anime style"
- Choice expansion: {red|blue|green} → "blue"
- LoRA syntax processing: <lora:base_effect:0.5> → applied to MODEL
- Result: populated_text = "anime style, blue, lora:base_effect:0.5"
Random LoRA Loader Processing:
- Auto-removes LoRA syntax from populated_text: "anime style, blue"
- Random LoRA selection: character_alice, concept_magic
- Final prompt: "anime style, blue, alice, blonde hair, magic circle"
- MODEL/CLIP: base_effect + character_alice + concept_magic (all applied)

Benefits

✅ Delegate wildcard functionality to Wildcard Encode (specialized tool)
✅ Delegate random LoRA selection to this node (simple configuration)
✅ Prompt information fully inherited
✅ LoRA syntax automatically cleaned (no noise)

3-Group Example

Group 1: style LoRAs
  - num_loras_1: 2
  - model_strength_1: "0.6-0.9"  ← Random (0.1 increments)
  - clip_strength_1: "0.6-0.9"   ← Random (0.1 increments)

Group 2: character LoRAs
  - num_loras_2: 1
  - model_strength_2: "1.0"      ← Fixed
  - clip_strength_2: "1.0"       ← Fixed

Group 3: Unused
  - lora_folder_path_3: (empty)
  - num_loras_3: 0

→ Result: 2 style LoRAs (variable strength) + 1 character LoRA (fixed strength) = 3 LoRAs total

Settings

Folder Scanning

All three nodes scan the folders you point them at the same way.

Supported extensions: .safetensors, .pt, .ckpt (case-insensitive).

Note: .pt and .ckpt files have no readable embedded metadata, so trigger words for them come only from .metadata.json / .info sidecar files.

Symbolic links are followed. A subfolder that is a symlink is descended into, matching ComfyUI's own folder scan — so LoRAs that only exist behind a link are included as candidates, just as they appear in ComfyUI's native LoRA dropdown. Link loops are detected and skipped rather than hanging the scan, and a folder reachable through two different links is only scanned once.

The candidate list is sorted before selection. This makes a given seed reproduce the same selection on any machine. Prior to v1.3.0 the order came straight from the filesystem, so the same seed could pick different LoRAs on a different PC.

⚠️ Because the ordering changed, an existing workflow may select a different set of LoRAs after upgrading to v1.3.0, even with a fixed seed. If a result you were relying on changes, re-roll or re-pick the seed.

Common Settings

Setting	Description	Default
`token_normalization`	Token normalization method	`none`
`weight_interpretation`	Prompt emphasis notation interpretation	`A1111`
`additional_prompt_positive`	Additional positive prompt (combined with trigger words)	(empty)
`additional_prompt_negative`	Additional negative prompt	(empty)
`trigger_word_source`	Trigger word source	`json_combined`
`seed`	Random selection seed	`0`

Important Notes on additional_prompt

Supported:

Regular prompt text
Text input from other nodes (e.g., Wildcard Encode's populated_text)

Not Supported (automatically removed):

<lora:xxx:0.8> format LoRA syntax
{a|b|c} format wildcard syntax
__filename__ format wildcard syntax

Why:

This node applies LoRAs via folder specification
Wildcard functionality should be handled by dedicated nodes (Wildcard Encode, etc.)
Writing LoRA syntax causes filename to become meaningful tokens as noise

Example (problematic case):

Input: "1girl, <lora:anime_style:0.8>, beautiful"
↓
CLIP tokenization: [1girl, lora, anime, style, 0, 8, beautiful]
                            ^^^^^ ^^^^^
                            Unintended words added to prompt ❌

Correct usage:

additional_prompt_positive: "1girl, beautiful"  ← No LoRA syntax ✅
LoRA application: Use folder specification feature ✅

Or, when connecting from Wildcard Encode, LoRA syntax is automatically removed:

[Wildcard Encode] populated_text: "1girl, <lora:style:0.8>, beautiful"
       ↓
[This Node] additional_prompt_positive received → LoRA syntax auto-removed
       ↓
Final prompt: "1girl, beautiful" ✅

Group Settings (1-3)

Each group can be configured individually:

Setting	Description	Default
`lora_folder_path_X`	LoRA folder absolute path	(empty)
`include_subfolders_X`	Include subfolders (symbolic links are followed)	`true`
`unique_by_filename_X`	Exclude duplicate filenames	`true`
`model_strength_X`	MODEL application strength	`"1.0"`
`clip_strength_X`	CLIP application strength	`"1.0"`
`num_loras_X`	Number of LoRAs to select	Group 1: `1`, Groups 2/3: `0`

Strength Specification (Important)

Strength fields accept fixed values or random ranges, including negative values.

Fixed Value

Input: "1.0"
→ Always applied at 1.0

Input: "0.55"
→ Rounded to 0.55 (2 decimals)

Input: "-0.5"
→ Negative LoRA at -0.5 ✅ (v1.2.0)

Random Range

Format: min-max

Input: "0.6-0.9"
→ Random from [0.6, 0.7, 0.8, 0.9] (0.1 increments)

Input: "-0.8--0.3"
→ Random from [-0.8, -0.7, -0.6, -0.5, -0.4, -0.3] ✅ (v1.2.0)

Input: "-0.5-0.5"
→ Random from [-0.5, -0.4, ..., 0.4, 0.5] ✅

v1.2.0 Improvements:

✅ Negative range support: -0.8--0.3 now works correctly
✅ 0.1 increments for all ranges: Easier to observe effect differences
✅ Fixed values rounded to 2 decimals: 0.847 → 0.85

Behavior

All Nodes:

Range: 0.1 increments list → random selection
Fixed: Rounded to 2 decimals

Example:

model_strength_1: "0.6-0.9"
→ Each execution randomly selects from [0.6, 0.7, 0.8, 0.9]
→ Console: "[RandomLoRALoader] Selected strength: 0.7"

MODEL Strength vs CLIP Strength

MODEL Strength

Affects: Image itself (art style, composition, colors, shapes)
Effect: Visual characteristics of generated image
Example:
- High value: Strong art style transfer
- Low value: Subtle style hints

CLIP Strength

Affects: Prompt interpretation (concept understanding, language-image association)
Effect: How strongly the model follows prompt instructions
Example:
- High value: Strong semantic guidance from trigger words
- Low value: Weak semantic influence

Recommended Settings

For most use cases:

model_strength: 0.8
clip_strength: 0.8
→ Balanced application

For subtle style application:

model_strength: 0.4-0.6
clip_strength: 0.8-1.0
→ Light visual effect, strong semantic guidance

For strong style override:

model_strength: 1.0-1.2
clip_strength: 0.6-0.8
→ Strong visual effect, moderate semantic guidance

Trigger Word Retrieval Methods

json_combined (Default)

Combines all trigger word patterns from JSON metadata.

Source:

activation text field
ss_tag_frequency keys (first 3 highest frequency)
modelspec.description keywords
ss_output_name keywords

Example:

{
  "activation text": "anime style, detailed eyes",
  "ss_tag_frequency": {
    "1girl": 500,
    "blue eyes": 300,
    "long hair": 250
  }
}

Result: "anime style, detailed eyes, 1girl, blue eyes, long hair"

json_random

Randomly selects one trigger word pattern.

Example:

Candidates:
- "anime style, detailed eyes"
- "1girl, blue eyes, long hair"
- "anime, girl, portrait"

Random selection: "1girl, blue eyes, long hair"

json_sample_prompt

Retrieves sample prompts from JSON metadata.

Source fields (priority order):

sample_prompts
civitai.images[0].meta.prompt

Processing:

Automatically removes LoRA syntax (<lora:xxx:0.8>)
Replaces with node-configured strength
Extracts negative prompt if available

Example:

{
  "sample_prompts": "1girl, <lora:style:0.9>, beautiful, detailed eyes"
}

Result: "1girl, beautiful, detailed eyes" (LoRA syntax removed)

metadata

Reads trigger words directly from embedded LoRA file metadata.

Source:

Safetensors metadata keys
Similar to json_combined but reads from LoRA file directly

Use when:

No external JSON files available
Want to use original training metadata

Outputs

All nodes provide the following outputs:

Output	Type	Description
`MODEL`	MODEL	Model with LoRAs applied
`CLIP`	CLIP	CLIP with LoRAs applied
`positive`	CONDITIONING	Positive conditioning (LoRA syntax removed)
`negative`	CONDITIONING	Negative conditioning
`positive_text`	STRING	Positive prompt text without LoRA syntax — exactly what is encoded into `positive` (v1.4.0)
`negative_text`	STRING	Full negative prompt text
`preview`	IMAGE	Batch of preview images for selected LoRAs (v1.1.0)
`lora_text`	STRING	Positive prompt text with LoRA syntax (the former `positive_text` content, v1.4.0)
`keyword`	STRING	Filtered nodes only. The keywords the selected LoRA actually matched (v1.5.0). See Keyword Output

⚠️ Output changes in v1.4.0

positive_text no longer contains <lora:...> tags; the previous content now comes from the new lora_text output (added at the end). All other outputs keep their positions and types, so existing connections stay valid.

positive_text fed a text encoder: no rewiring needed. The tags are no longer read as text

positive_text was used to record the LoRA syntax (saved prompts, metadata): reconnect that to lora_text

The Filtered nodes also now tidy the stray , , left behind after removing tags (and leading/trailing commas) before encoding, as the 3-folder node already did. Because the encoded text changes, results with the same seed may differ from earlier versions. If you need identical results to an earlier version, keep using that version.

Text Output Format

lora_text example:

<lora:style_anime:0.8:0.8>, <lora:character_alice:1.0:1.0>, anime style, alice, blonde hair, 1girl, beautiful

LBW node lora_text example:

<lora:style_anime:0.8:0.8:lbw=1,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1>, anime style, 1girl, beautiful

positive_text / positive (CONDITIONING) content:

anime style, alice, blonde hair, 1girl, beautiful
(LoRA syntax removed for clean prompt)

Preview Images (v1.1.0)

Supported formats (priority order):

Static images (.png, .jpg, .jpeg) - Always works
Animated images (.gif, .webp) - First frame, always works
Video files (.mp4, .webm, .avi, .mov) - First frame, requires opencv-python

File matching:

Files starting with the LoRA filename (case-insensitive)
Example: style_anime.safetensors matches:
- style_anime.png ✅
- style_anime_preview.jpg ✅
- STYLE_ANIME.PNG ✅

Filtered Random LoRA Loader

Single folder with keyword filtering.

Key Features

Keyword Filtering: Filter LoRAs by space-separated keywords
AND/OR/OFF Modes: Flexible filtering logic (OFF: v1.5.0)
Phrase Matching: Use quotes for exact phrases
Metadata Search: Search in filenames or embedded metadata
Fast Caching: Instant metadata access after first load
Serial Connection Ready: Chain multiple instances for complex workflows

Parameters

Parameter	Description	Default
`lora_folder_path`	LoRA folder path	(empty)
`keyword_filter`	Space-separated keywords or quoted phrases	(empty)
`filter_mode`	`AND` / `OR` / `OFF`	`AND`
`search_in_metadata`	Search in JSON/embedded metadata	`false`
`num_loras`	Number of LoRAs to select	`1`
`model_strength`	MODEL strength (fixed or range)	`"1.0"`
`clip_strength`	CLIP strength (fixed or range)	`"1.0"`
`include_subfolders`	Include subfolders (symbolic links are followed)	`true`
`unique_by_filename`	Exclude duplicate filenames	`true`

Keyword Filter Syntax

Basic keywords (AND mode):

keyword_filter: "anime girl"
→ Matches files containing BOTH "anime" AND "girl"

OR mode:

filter_mode: OR
keyword_filter: "anime realistic"
→ Matches files containing "anime" OR "realistic"

OFF mode (v1.5.0):

filter_mode: OFF
keyword_filter: "anime realistic"
→ No filtering; picks from every LoRA in the folder
→ The keywords you typed stay, so you can switch the filter off for a while without deleting them

Phrase matching:

keyword_filter: '"anime style" detailed'
→ Must contain exact phrase "anime style" AND word "detailed"

Multiple phrases:

keyword_filter: '"anime style" "detailed eyes" red'
→ Must contain "anime style" AND "detailed eyes" AND "red"

Keyword Output (v1.5.0)

Outputs the keywords the selected LoRA actually matched, joined with _. The LBW node works the same way.

`keyword_filter`	`filter_mode`	Selected LoRA (filename)	`keyword` output
`aaa bbb`	AND	`aaa_bbb_xxx`	`aaa_bbb`
`aaa bbb`	OR	`aaa_xxx`	`aaa`
`aaa bbb`	OR	`bbb_xxx`	`bbb`
`"aaa bbb"`	AND / OR	`aaa bbb xxx`	`aaa_bbb`
`"aaa bbb" ccc`	AND	`aaa bbb ccc`	`aaa_bbb_ccc`
(empty)	AND / OR	any	(empty)
(anything)	OFF	any	(empty)

Spaces inside a phrase also become _
In OR mode, if the selected LoRA contains several of the keywords, or num_loras is 2 or more, every matched keyword is joined in input order (e.g. aaa_bbb)
Keywords are output with the case you typed (matching itself ignores case)
With search_in_metadata ON, keywords matched in the metadata are included too
Empty when filter_mode is OFF, when keyword_filter is empty, when no LoRA matched, or when num_loras is 0

keyword was added as the last (ninth) output, so existing workflow connections keep working.

Metadata Search

Filename search (default):

search_in_metadata: false
→ Fast, searches only filenames

Metadata search:

search_in_metadata: true
→ Slower, searches JSON/embedded metadata
→ Cached after first search (instant on 2nd+)

Performance:

Initial load (SSD):
- 1,000 files: ~2 seconds
- 5,000 files: ~10 seconds
- 10,000 files: ~20 seconds
2nd+ time: Instant (<100ms)
Memory: ~150MB per 10,000 files

Use Cases

Example 1: Character Selection

lora_folder_path: "/path/to/all_loras"
keyword_filter: "character girl"
filter_mode: AND
num_loras: 1

Result: Select 1 character LoRA containing both "character" and "girl"

Example 2: Multiple Instances (Recommended)

[Load Checkpoint]
  ↓
[Filtered Random LoRA Loader #1]
  keyword_filter: "style anime"
  num_loras: 1
  ↓
[Filtered Random LoRA Loader #2]
  keyword_filter: "character"
  num_loras: 1
  ↓
[KSampler]

Benefits:

Different filtering per category
Independent strength settings
More flexibility

Example 3: Metadata Search

search_in_metadata: true
keyword_filter: "detailed eyes"

Searches in:

Filenames
JSON activation text
JSON ss_tag_frequency
JSON modelspec.description
Embedded LoRA metadata

Filtered Random LoRA Loader (LBW) 🆕 NEW v1.2.0

Advanced node with LoRA Block Weight support.

What is LBW?

LoRA Block Weight allows you to control which parts of the U-Net are affected by the LoRA:

INPUT blocks  → Structure, composition, layout
   ↓
MIDDLE block  → Overall features
   ↓
OUTPUT blocks → Style, details, fine-tuning

Key Features

4 Preset Modes:
- Style Focused: OUTPUT blocks only (style without changing structure)
- Character Focused: Balanced IN+MID+OUT (character features)
- Structure/Composition Only: INPUT+MID blocks (layout without style)
- Balanced / Soft: Gentle application (natural results)
Preset: Random: Randomly select one of the 4 presets
Direct Input: Custom weight specification
Automatic SD1.5/SDXL Detection: No manual model type selection
Automatic Weight Adjustment: Handles mismatched element counts

Parameters

Same as Filtered Random LoRA Loader, plus:

Parameter	Description	Default
`weight_mode`	LBW preset or Direct Input	`Normal (All 1.0)`
`lbw_input`	Custom weights (for Direct Input)	(empty)

Weight Modes

Normal (All 1.0)

Standard LoRA application without block weighting.

Style Focused

SDXL: 1,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1
SD1.5: 1,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1

OUTPUT blocks only
Style without changing structure
Use for: Art style LoRAs, effect LoRAs

Character Focused

SDXL: 1,1,1,1,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1
SD1.5: 1,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1

Balanced IN+MID+OUT
Character features preserved
Use for: Character LoRAs, person LoRAs

Structure/Composition Only

SDXL: 1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0
SD1.5: 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0

INPUT+MID blocks only
Structure without changing style
Use for: Pose LoRAs, composition LoRAs

Balanced / Soft

SDXL: 1,1,1,1,0,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0
SD1.5: 1,1,1,1,0,0,0,1,1,1,1,1,0,0,0,0,0

Gentle application
Natural results
Use for: General LoRAs, multi-LoRA workflows

Preset: Random

Randomly selects one of the 4 presets above each execution.

Direct Input

weight_mode: Direct Input
lbw_input: "1,0.5,0.5,0,0,0,0,0,0,0,1,0.8,0.8,0.8,0.8,0.8,0.5,0.5,0.5,0.5"

Custom weight specification
SDXL: 20 elements, SD1.5: 17 elements
Auto-adjusted if count doesn't match

Use Cases

Example 1: Style Only

[Filtered Random LoRA Loader (LBW)]
  keyword_filter: "anime watercolor"
  weight_mode: Style Focused
  ↓
[KSampler]

Result: Applies anime watercolor style without changing composition

Example 2: Structure then Style

[Load Checkpoint]
  ↓
[Filtered Random LoRA Loader (LBW)]
  keyword_filter: "pose"
  weight_mode: Structure/Composition Only
  ↓
[Filtered Random LoRA Loader (LBW)]
  keyword_filter: "oil painting"
  weight_mode: Style Focused
  ↓
[KSampler]

Result: Pose adjustment → Oil painting style applied

Example 3: Custom Weights

weight_mode: Direct Input
lbw_input: "1,0.5,0.5,0.5,0,0,0,0,0,0,1,1,1,1,1,1,0.5,0.5,0.5,0.5"

Result: Custom block weight distribution

Advanced: Custom Presets

You can edit presets directly in the source file:

File: filtered_random_lora_loader_lbw.py (lines 29-41)

See README_LBW.md for detailed customization guide.

Tips & Tricks

Serial Connection for Multiple Effects

Connect nodes in series to apply different LoRA categories:

[Load Checkpoint]
  ↓
[Filtered Random LoRA Loader (LBW)] (Structure/Composition Only)
  keyword_filter: "pose"
  ↓
[Filtered Random LoRA Loader (LBW)] (Style Focused)
  keyword_filter: "watercolor"
  ↓
[KSampler]

Benefits:

Different settings per category
Cumulative LoRA effects
More control over final result

Strength Randomization for Variety

Use ranges instead of fixed values:

model_strength_1: "0.6-0.9"
clip_strength_1: "0.6-0.9"

Effect:

Each generation uses different strength
More variety in results
Easier to find optimal strength

Keyword Filtering Best Practices

Use specific keywords:

✅ Good: "anime girl"
❌ Too broad: "anime"

Combine with metadata search:

search_in_metadata: true
keyword_filter: "detailed eyes"

Use phrases for precision:

keyword_filter: '"anime style" detailed'
→ Must contain exact phrase "anime style" AND word "detailed"

Wildcard Encode + Random LoRA Loader

Ultimate dynamic workflow:

[Wildcard Encode (Inspire)]
  text: "__style__, __pose__, {red|blue|green}"
  ↓
  populated_text ──→ [Random LoRA Loader]
                     additional_prompt_positive
  ↓
[KSampler]

Result:

Wildcard expansion for prompts
Random LoRA selection for variety
Clean prompt without LoRA syntax
Maximum flexibility

Preview Images for Organization

Enable preview display to:

See which LoRAs were selected
Verify LoRA appearance before generation
Organize LoRA collections visually

Recommended workflow:

[Random LoRA Loader]
  preview ──→ [Preview Image] (for verification)
  MODEL ──→ [KSampler]

Troubleshooting

No LoRAs Found

Check:

LoRA folder path is correct (absolute path)
.safetensors, .pt or .ckpt files exist in the folder
include_subfolders setting if LoRAs are in subfolders
Console for error messages

Example valid paths:

Windows: C:/ComfyUI/models/loras/style
Linux/Mac: /home/user/ComfyUI/models/loras/style

No Trigger Words Found

Check:

Metadata files (.metadata.json or .info) exist
JSON files contain required fields:
- activation text
- ss_tag_frequency
- sample_prompts
- civitai.trainedWords or civitai.images
Try different trigger_word_source settings

Verify JSON format:

{
  "activation text": "anime style, detailed",
  "ss_tag_frequency": {...},
  "modelspec.description": "...",
  "sample_prompts": "..."
}

LoRA Not Applied

Check:

num_loras is not 0
Strength values are not all 0
MODEL/CLIP outputs are connected properly
Console for LoRA application messages
For LBW node: weight_mode is not "Normal (All 1.0)" with all weights set to 0

Preview Images Not Showing

Check:

Preview files match LoRA filename
For video files: opencv-python is installed (pip install opencv-python)
Supported formats: .png, .jpg, .jpeg, .gif, .webp, .mp4, .webm, .avi, .mov
File permissions are correct

Without opencv-python:

Static images work ✅
Animated images work ✅
Video files show black screen ⚠️

Strength Not Working with Negative Values (Fixed in v1.2.0)

Previous issue (v1.1.0):

Input: "-0.5"
Result: 1.0 ❌ (converted to default)

Fixed in v1.2.0:

Input: "-0.5"
Result: -0.5 ✅ (negative LoRA applied correctly)

Input: "-0.8--0.3"
Result: Random from [-0.8, -0.7, -0.6, -0.5, -0.4, -0.3] ✅

Wildcard Encode Issues

If wildcard text appears in output:

Check connection: Connect Wildcard Encode's populated_text to additional_prompt_positive
Wrong output: Don't connect text (input) to this node
LoRA syntax appearing: This node automatically removes it

Example workflow:

[Wildcard Encode (Inspire)]
  populated_text ──→ [Random LoRA Loader]
  (not "text")       additional_prompt_positive

Duplicate LoRAs Selected

Solution: Enable unique_by_filename

unique_by_filename_1: true

This prevents selecting the same LoRA from different subfolders:

/style/lora.safetensors
/backup/lora.safetensors
→ Only one selected ✅

Detailed Documentation

README_LBW.md - Complete LBW documentation (English)
README_LBW_ja.md - LBW詳細ガイド（日本語）
README_ja.md - 日本語メインドキュメント

Disclaimer and Support Policy

Disclaimer

This node is provided as-is with no technical support
No warranty or guarantee of functionality
No guaranteed compatibility with future ComfyUI updates
Bug reports and feature requests may not be addressed
Use at your own risk

Support Status

❌ No individual support via issues or email
❌ No guaranteed bug fixes or feature additions
✅ Code is open source - feel free to fork and modify
✅ Community discussions welcome (no promises of response)

Reporting Issues

While support is not guaranteed, you can:

Check existing issues in the repository
Review this README and troubleshooting section
Open an issue (may or may not be addressed)
Fork and fix it yourself

License

MIT License

Changelog

v1.2.0 (2026-01-13)

Added

✅ NEW NODE: Filtered Random LoRA Loader (LBW)
✅ LoRA Block Weight (LBW) support with 4 presets
✅ Automatic SD1.5/SDXL detection for LBW
✅ Preset: Random mode for LBW
✅ Direct Input mode for custom LBW weights
✅ Automatic weight adjustment for LBW
✅ Negative value support for strength: -0.5, -0.8--0.3 now work correctly
✅ Video preview support: .mp4, .webm, .avi, .mov (requires opencv-python)
✅ Strength precision improvements:
- Range specification: 0.1 increments for all nodes
- Fixed values: Rounded to 2 decimals

Fixed

✅ Critical bug: Negative single values ("-0.5") now work correctly (previously converted to 1.0)
✅ Critical bug: Negative ranges ("-0.8--0.3") now work correctly (previously caused errors)
✅ Improved strength parsing with regex pattern matching

Changed

✅ Unified strength behavior across all 3 nodes
✅ Renamed node: "Filtered Random LoRA Loader (Advanced)" → "Filtered Random LoRA Loader (LBW)"
✅ Enhanced documentation with model compatibility warnings

v1.1.0 (2026-01-04)

Added

✅ NEW NODE: Filtered Random LoRA Loader
✅ Keyword filtering with AND/OR modes
✅ Metadata search with caching
✅ Preview image output (IMAGE type)
✅ Duplicate filename handling (unique_by_filename)
✅ Animated image support (.gif, .webp)

Changed

✅ Breaking: Split additional_prompt into additional_prompt_positive and additional_prompt_negative
✅ Unified strength precision to 1 decimal place
✅ Improved LoRA syntax removal

Fixed

✅ Fixed negative_text output in json_sample_prompt mode
✅ Fixed LoRA syntax appearing in metadata trigger words
✅ Fixed duplicate parameter in INPUT_TYPES

v1.0.0 (2025-12-30)

Added

✅ Initial release
✅ Random LoRA Loader with 3-group support
✅ Multi-source metadata reading
✅ Strength randomization with range specification
✅ Trigger word extraction
✅ Wildcard Encode compatibility
✅ CONDITIONING output with cleaned prompts

References

Enjoy flexible LoRA randomization with precise control! 🎲✨

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
images		images
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
LICENSE		LICENSE
README.md		README.md
README_LBW.md		README_LBW.md
README_LBW_ja.md		README_LBW_ja.md
README_ja.md		README_ja.md
__init__.py		__init__.py
filtered_random_lora_loader.py		filtered_random_lora_loader.py
filtered_random_lora_loader_lbw.py		filtered_random_lora_loader_lbw.py
pyproject.toml		pyproject.toml
random_lora_loader.py		random_lora_loader.py

Latest commit

History

Folders and files

Repository files navigation

Random LoRA Loader for ComfyUI

⚠️ Important: Supported Models

Nodes Overview

Random LoRA Loader (Original)

Filtered Random LoRA Loader

Filtered Random LoRA Loader (LBW) 🆕 NEW v1.2.0

Key Features

Common Features (All Nodes)

Filtered Nodes (Filtered & LBW)

LBW Node Only 🆕

Installation

⚠️ Optional: Video Preview Support

Requirements

Steps

Uninstallation

Usage

Basic Usage (Random LoRA Loader)

Wildcard Encode Integration (Recommended)

Recommended Connection

How It Works

Benefits

3-Group Example

Settings

Folder Scanning

Common Settings

Important Notes on additional_prompt

Group Settings (1-3)

Strength Specification (Important)

Fixed Value

Random Range

Behavior

MODEL Strength vs CLIP Strength

MODEL Strength

CLIP Strength

Recommended Settings

Trigger Word Retrieval Methods

json_combined (Default)

json_random

json_sample_prompt

metadata

Outputs

Text Output Format

Preview Images (v1.1.0)

Filtered Random LoRA Loader

Key Features

Parameters

Keyword Filter Syntax

Keyword Output (v1.5.0)

Metadata Search

Use Cases

Example 1: Character Selection

Example 2: Multiple Instances (Recommended)

Example 3: Metadata Search

Filtered Random LoRA Loader (LBW) 🆕 NEW v1.2.0

What is LBW?

Key Features

Parameters

Weight Modes

Normal (All 1.0)

Style Focused

Character Focused

Structure/Composition Only

Balanced / Soft

Preset: Random

Direct Input

Use Cases

Example 1: Style Only

Example 2: Structure then Style

Example 3: Custom Weights

Advanced: Custom Presets

Tips & Tricks

Serial Connection for Multiple Effects

Strength Randomization for Variety

Keyword Filtering Best Practices

Wildcard Encode + Random LoRA Loader

Preview Images for Organization