nodes/shin131002-RandomLoRALoader

Fork 0

Files

T

shin132002 6e1cf9b518 v1.5.0: Filtered nodes: add keyword output and filter_mode OFF

2026-09-29 01:24:30 +09:00

15 KiB

Raw Permalink Blame History

Filtered Random LoRA Loader (LBW)

LoRA Block Weight (LBW) support with automatic SD1.5/SDXL detection

Overview

Filtered Random LoRA Loader (LBW) is the Filtered Random LoRA Loader with LoRA Block Weight (LBW) support added. It lets you control precisely which parts of the U-Net a LoRA is applied to.

Features

Core Features (shared with Filtered Random LoRA Loader)

✅ Random LoRA selection from a single folder
✅ Keyword filtering (AND/OR/OFF modes)
✅ Output of the matched keywords (keyword, v1.5.0)
✅ Metadata search (filename or embedded metadata)
✅ Preview images (static/animated/video support)
✅ Trigger word extraction
✅ Duplicate removal by filename
✅ Direct serial connection support

Advanced Features (LBW)

✅ LoRA Block Weight (LBW) support: set a separate strength for each U-Net block to control style, character, composition and so on independently
✅ Automatic SD1.5/SDXL detection
✅ 4 preset modes + custom input
✅ Automatic weight adjustment

LoRA Block Weight (LBW)

What is LBW?

LoRA Block Weight lets you control which parts of the U-Net the LoRA affects. The Stable Diffusion U-Net is made up of three parts:

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

By adjusting the weight of each block, you can precisely control the LoRA's effect.

Weight Modes

1. Normal (All 1.0) - Default

All blocks weighted at 1.0 (standard LoRA application, no LBW).

2. Style Focused

Enables OUTPUT blocks only.

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

Best for:

Style LoRAs (anime style, watercolor, etc.)
Art style and effect LoRAs
Changing the visual aesthetic

Effect:

✅ Strong style changes
❌ Minimal structure/character changes (ideal when you want to keep the existing composition)

3. Character Focused

Enables INPUT (early) + MIDDLE + OUTPUT blocks.

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

Best for:

Character and person LoRAs
Concept LoRAs
Preserving faces/features

Effect:

✅ Strong character features
✅ Balanced overall effect

4. Structure/Composition Only

Enables INPUT + MIDDLE blocks 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

Best for:

Pose LoRAs
Composition and layout LoRAs

Effect:

✅ Strong structure/pose changes
❌ Minimal style changes

5. Balanced / Soft

Soft application across selected blocks. The later OUTPUT blocks are turned down to avoid over-application.

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

Best for:

Subtle effects and fine adjustment
Multi-LoRA workflows
General-purpose LoRAs

Effect:

✅ Balanced, gentle application with a natural finish
✅ Good for stacking multiple LoRAs

6. Preset: Random

Randomly picks one of the 4 presets above on each run.

weight_mode: "Preset: Random"
→ One of the 4 presets is picked on each run

Best for:

Experimentation
Variation generation
Random effects

7. Direct Input

Custom weight specification.

Format:

SDXL: 20 comma-separated values
SD1.5: 17 comma-separated values

Example (SDXL):

weight_mode: "Direct Input"
lbw_input: "1,0.5,0.5,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.7,0.7,0.7,1,1,1,1,1,1,1"

Auto-adjustment:

If too few elements: pads with 1.0 at the end
If too many elements: truncates from the end

Block Structure

SDXL (20 elements)

BASE, IN00-IN08 (9), MID (1), OUT00-OUT08 (9)

Index	Block	Role
0	BASE	Overall base
1-9	INPUT (IN00-08)	Structure/composition
10	MIDDLE (M00)	Intermediate features
11-19	OUTPUT (OUT00-08)	Style/details

Weight example:

1,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1
BASE,IN00-08 (OFF),M00,OUT00-08 (ON)
→ LoRA applied to OUTPUT blocks only (style focused)

SD1.5 (17 elements)

BASE, IN01,IN02,IN04,IN05,IN07,IN08 (6), MID (1), OUT03-OUT11 (9)

Index	Block	Role
0	BASE	Overall base
1-6	INPUT (IN01,02,04,05,07,08)	Structure/composition
7	MIDDLE (M00)	Intermediate features
8-16	OUTPUT (OUT03-11)	Style/details

Weight example:

1,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1
BASE,IN (OFF),M00,OUT (ON)
→ LoRA applied to OUTPUT blocks only (style focused)

Automatic Detection

LoRA Type Detection

The node automatically detects whether a LoRA is SD1.5 or SDXL by analyzing the model structure:

SD1.5: output_blocks_0 to output_blocks_11 (12 blocks)
SDXL: output_blocks_0 to output_blocks_8 (9 blocks)

SDXL detected  → 20-element weights are used
SD1.5 detected → 17-element weights are used

You don't need to specify the model type manually! ✅

Weight Adjustment

If the number of weights doesn't match the detected LoRA type, it is adjusted automatically:

Example 1: Too few weights

Input: 1,0,0 (3 elements)
Detected: SDXL (requires 20)
→ Auto-adjusted: 1,0,0,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1

Example 2: Too many weights

Input: 1,0,0,...,0,0,0 (25 elements)
Detected: SDXL (requires 20)
→ Auto-adjusted: 1,0,0,...,0,0,0 (first 20 only)

Usage

Basic Workflow

[Filtered Random LoRA Loader (LBW)]
  MODEL → [KSampler]
  CLIP → [CLIP Text Encode]
  lora_text → [Show Text] (optional, for debugging)

Steps:

Place the node
```
[Filtered Random LoRA Loader (LBW)]
```

Set the folder and keywords

lora_folder_path: "path/to/loras"
keyword_filter: "anime style"
num_loras: 1

Choose the LBW mode
```
weight_mode: "Style Focused"
```

Connect MODEL/CLIP

MODEL → [KSampler]
CLIP → [CLIP Text Encode]

Parameters

LoRA Settings

lora_folder_path: Path to LoRA folder
num_loras: Number of LoRAs to apply (0-20)
model_strength: LoRA strength for MODEL (e.g., "1.0" or "0.6-0.9")
clip_strength: LoRA strength for CLIP (e.g., "1.0" or "0.6-0.9")

Folder scanning behaviour (supported extensions, symbolic links, sorted candidate order) is shared with the other nodes — see Folder Scanning in the main README.

Keyword Filter

keyword_filter: Space-separated keywords (e.g., style anime or "anime style" red)
filter_mode: AND / OR / OFF (OFF turns the filter off but keeps the keywords, v1.5.0)
search_in_metadata: Search in JSON/embedded metadata (slower)

LBW Settings

weight_mode: Normal / 4 Presets / Random / Direct Input
lbw_input: Custom weights (only for Direct Input mode)

Other

trigger_word_source: Trigger word extraction method
seed: Random seed

Output Format

LBW Syntax

When LBW is applied, the lora_text output (positive_text before v1.4.0) includes LBW syntax:

Normal (All 1.0):

<lora:style_anime:0.8:0.8>

With LBW (Style Focused, SDXL):

<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>

Note: The LBW is applied internally to the MODEL. The syntax in lora_text is for reference/re-use with other nodes.

Keyword Output (v1.5.0)

Outputs the keywords the selected LoRA actually matched, joined with _ (last output). AND mode gives all keywords (aaa bbb → aaa_bbb); OR mode gives the one that matched (aaa or bbb). Spaces inside a phrase also become _ ("aaa bbb" ccc → aaa_bbb_ccc). Empty when filter_mode is OFF. See Keyword Output in README.md for details.

Examples

Example 1: Change the style only

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

Result: The composition stays the same; only the style becomes anime-like

Example 2: Staged application with multiple LoRAs

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

Result: Pose adjustment → watercolor style applied

Example 3: Character + style

[Load Checkpoint]
  ↓
[Filtered Random LoRA Loader (LBW)] (Character Focused)
  keyword_filter: "character_name"
  ↓
[Filtered Random LoRA Loader (LBW)] (Style Focused)
  keyword_filter: "art_style"
  ↓
[KSampler]

Result: Character features preserved + style changed

Tips

1. Preset Selection Guide

LoRA Type	Recommended Preset	Why
Style/Art	Style Focused	OUTPUT blocks control style
Character	Character Focused	Balanced IN+MID+OUT preserves features
Pose/Composition	Structure/Composition Only	INPUT blocks control layout
General/fine adjustment	Balanced / Soft	Gentle, stackable effect

2. Effect Visibility

The difference between presets is more visible with:

✅ Strong LoRAs (high model_strength)
✅ Single LoRA application (num_loras=1)
✅ Style or concept LoRAs

The difference is less visible with:

⚠️ Weak LoRAs (low model_strength)
⚠️ Multiple LoRAs applied simultaneously
⚠️ LoRAs whose training effect is spread evenly across blocks

3. Zero Weight = No Effect

Setting all weights to 0 completely disables the LoRA effect:

0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0

This is equivalent to not applying the LoRA at all.

Weight guide:

0: Completely off
0.5: Weaker
1.0: Standard
1.5 or more: Stronger (watch for over-application)

4. Serial Connection

For applying different LBW settings to different LoRAs, connect nodes in series:

[Filtered Random LoRA Loader (LBW)] (Style Focused)
  ↓
[Filtered Random LoRA Loader (LBW)] (Character Focused)
  ↓
[KSampler]

All LoRAs within a single node instance use the same LBW setting.

Tips:

Structure → style is the basic order
Strength can also be adjusted per node

5. Troubleshooting

If there is no effect:

Check that the weights aren't all 0
Check that the LoRA is loaded correctly
Check the console for the LBW application message

If the effect is too strong:

Lower model_strength
Try the Balanced / Soft preset
Adjust the weights to 0.5–0.8

Requirements

⚠️ Supported Models

This node is 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 feature specifically targets SD1.5/SDXL U-Net architecture. Other model types will not work correctly.

Dependencies

ComfyUI
Pillow (included with ComfyUI)
torch (included with ComfyUI)

⚠️ Optional: Video Preview

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

pip install opencv-python

Without opencv-python:

✅ Static/animated images work
❌ Video files show black screen

Preview Images

Supports the following formats (in priority order):

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

The node automatically searches for preview files matching the LoRA filename.

File matching:

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

Advanced: Custom Presets

To create your own presets or change the existing ones, you can edit the source file directly.

How to Edit

File location:

ComfyUI/custom_nodes/RandomLoRALoader/filtered_random_lora_loader_lbw.py

Lines to edit: 29–41 (SDXL_PRESETS for SDXL, SD15_PRESETS for SD1.5)

Example 1: Add your own preset

SDXL_PRESETS = {
    "Style Focused": "1,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1",
    "Character Focused": "1,1,1,1,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1",
    "Structure/Composition Only": "1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0",
    "Balanced / Soft": "1,1,1,1,0,0,0,0,0,0,1,1,1,1,1,1,0,0,0,0",
    # ↓ Your own preset
    "My Custom Mix": "0.5,0.5,0.5,0.5,0,0,0,0,0,0,1,0.8,0.8,0.8,0.8,0.8,0.5,0.5,0.5,0.5"
}

Example 2: Adjust an existing preset

SDXL_PRESETS = {
    # Make Style Focused stronger
    "Style Focused": "1,0,0,0,0,0,0,0,0,0,0,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2",
    
    # Make Character Focused milder
    "Character Focused": "0.8,0.8,0.8,0.8,0,0,0,0,0,0,0.8,0.8,0.8,0.8,0.8,0.8,0.8,0.8,0.8,0.8",
    ...
}

Example 3: Rename presets

SDXL_PRESETS = {
    "Style Only": "1,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1",
    "Character First": "1,1,1,1,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1,1",
    "Composition Only": "1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0",
    ...
}

Important Notes

⚠️ Before editing:

Make a backup of the original file
Watch the Python syntax (commas, quotes)

⚠️ Number of weights:

SDXL presets: exactly 20 elements
SD1.5 presets: exactly 17 elements
Mismatched counts are auto-adjusted, but specifying them exactly is more reliable

⚠️ After editing:

Fully restart ComfyUI to apply the changes
Check the console for syntax errors

💡 Tip: If editing the file is a hassle, use the "Direct Input" mode!

Disclaimer

No technical support provided
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

Changelog

v1.2.0 (2026-01-13)

✅ Added LoRA Block Weight (LBW) support
✅ Automatic SD1.5/SDXL detection
✅ 4 preset modes + custom input
✅ Automatic weight adjustment
✅ Video preview support (.mp4, .webm, etc.)

v1.1.0 (2026-01-04)

✅ Added Filtered Random LoRA Loader
✅ Keyword filtering with AND/OR modes
✅ Metadata search support
✅ Preview image display

v1.0.0 (2025-12-30)

✅ Initial release
✅ Random LoRA Loader (3 groups)

License

MIT License

Author

konohana

Enjoy precise LoRA control with Block Weights! 🎨✨

15 KiB Raw Permalink Blame History Unescape Escape

Filtered Random LoRA Loader (LBW)

Overview

Features

Core Features (shared with Filtered Random LoRA Loader)

Advanced Features (LBW)

LoRA Block Weight (LBW)

What is LBW?

Weight Modes

1. Normal (All 1.0) - Default

2. Style Focused

3. Character Focused

4. Structure/Composition Only

5. Balanced / Soft

6. Preset: Random

7. Direct Input

Block Structure

SDXL (20 elements)

SD1.5 (17 elements)

Automatic Detection

LoRA Type Detection

Weight Adjustment

Usage

Basic Workflow

Parameters

LoRA Settings

Keyword Filter

LBW Settings

Other

Output Format

LBW Syntax

Keyword Output (v1.5.0)

Examples

Example 1: Change the style only

Example 2: Staged application with multiple LoRAs

Example 3: Character + style

Tips

1. Preset Selection Guide

2. Effect Visibility

3. Zero Weight = No Effect

4. Serial Connection

5. Troubleshooting

Requirements

⚠️ Supported Models

Dependencies

⚠️ Optional: Video Preview

Preview Images

Advanced: Custom Presets

How to Edit

Example 1: Add your own preset

Example 2: Adjust an existing preset

Example 3: Rename presets

Important Notes

Disclaimer

Changelog

v1.2.0 (2026-01-13)

v1.1.0 (2026-01-04)

v1.0.0 (2025-12-30)

License

Author

15 KiB

Raw Permalink Blame History