ComfyAssets-ComfyUI-KikoTools/examples/documentation/lora_folder_batch.md

LoRA Folder Batch

Overview

The LoRA Folder Batch node automates the process of testing multiple LoRA models from a folder. This tool was adapted from comfyui-essentials-nodes (now in maintenance mode) and enhanced with batch processing capabilities for efficient LoRA evaluation workflows.

Attribution

This node is based on work from comfyui-essentials-nodes by cubiq. The original project is in maintenance-only mode, and we've adopted and enhanced these tools to ensure continued support and compatibility with modern ComfyUI workflows.

Features

• Automatic Folder Scanning: Discovers all .safetensors files in specified folders
• Natural Sorting: Intelligently sorts epochs (e.g., epoch_004, epoch_020, epoch_100)
• Pattern Filtering: Include/exclude LoRAs using regex patterns
• Flexible Strength Control: Single, multiple, or range-based strength values
• Batch Modes: Sequential or combinatorial strength application
• Epoch Detection: Automatically extracts epoch numbers from filenames
• Auto-Batching: Automatically splits large LoRA collections into manageable chunks to prevent UI disconnection

Node Properties

• Category: ComfyAssets/🧰 xyz-helpers
• Node Name: LoRAFolderBatch
• Function: batch_loras

Inputs

Required

Parameter	Type	Default	Description
folder_path	STRING	"."	Folder path relative to models/loras (or absolute)
strength	STRING	"1.0"	Strength values (see formats below)
batch_mode	DROPDOWN	sequential	[sequential, combinatorial] processing mode

Optional

Parameter	Type	Default	Description
include_pattern	STRING	""	Regex pattern to include files
exclude_pattern	STRING	""	Regex pattern to exclude files
max_loras	INT	50	Maximum LoRAs to process (when auto_batch disabled)
sort_order	DROPDOWN	natural	Sorting method [natural, alphabetical, newest, oldest]
auto_batch	DROPDOWN	disabled	Enable auto-batching for large collections [disabled, enabled]
batch_size	INT	25	Number of LoRAs per batch when auto-batching
batch_index	INT	0	Which batch to output (0-based) when auto-batching

Strength Format Options

• Single: "1.0" - Apply same strength to all LoRAs
• Multiple: "0.5, 0.75, 1.0" - Comma-separated values
• Range: "0.5...1.0+0.25" - Start...End+Step format

Outputs

Name	Type	Description
lora_params	LORA_PARAMS	Batch parameters for processing
lora_list	STRING	List of discovered LoRAs with epoch info
lora_count	INT	Number of LoRAs found

Usage Examples

Test All Epochs of a LoRA

LoRAFolderBatch → FluxSamplerParams → KSampler
    folder_path: "my_lora_training"
    strength: "1.0"
    batch_mode: sequential

Strength Testing for Each LoRA

LoRAFolderBatch → KSampler → Image Grid
    folder_path: "test_loras"
    strength: "0.5, 0.75, 1.0"
    batch_mode: combinatorial

Filter Specific Epochs

LoRAFolderBatch → Processing Pipeline
    folder_path: "training_results"
    include_pattern: "epoch_0[2-5]0"
    strength: "0.8...1.2+0.1"

Auto-Batch Large Collections

LoRAFolderBatch → FluxSamplerParams → KSampler
    folder_path: "massive_lora_collection"  # 100+ files
    strength: "1.0"
    auto_batch: enabled
    batch_size: 25
    batch_index: 0  # Change to 1, 2, 3... for subsequent batches

Batch Modes Explained

Sequential Mode

Each LoRA gets one strength value in order:

• LoRA1 → strength[0]
• LoRA2 → strength[1]
• LoRA3 → strength[0] (cycles if fewer strengths than LoRAs)

Combinatorial Mode

Each LoRA is tested with ALL strength values:

• LoRA1 → [0.5, 0.75, 1.0]
• LoRA2 → [0.5, 0.75, 1.0]
• LoRA3 → [0.5, 0.75, 1.0]

Auto-Batching for Large Collections

Overview

When testing large numbers of LoRAs (e.g., 75+ files), ComfyUI can experience UI disconnections or memory issues. Auto-batching solves this by automatically splitting your LoRA collection into smaller, manageable chunks.

How It Works

1. Enable Auto-Batching: Set auto_batch to "enabled"
2. Set Batch Size: Configure batch_size (default: 25, range: 5-100)
3. Select Batch: Use batch_index to choose which batch to process

Example: Testing 75 LoRAs

With 75 LoRAs and batch_size=25, the system creates 3 batches:

• Batch 0: LoRAs 1-25 (set batch_index=0)
• Batch 1: LoRAs 26-50 (set batch_index=1)
• Batch 2: LoRAs 51-75 (set batch_index=2)

Run your workflow 3 times, changing only the batch_index each time.

Visual Feedback

When auto-batching is enabled, the lora_list output includes batch information:

=== Batch 1/3 (LoRAs 1-25) ===

style-epoch-001
style-epoch-002
...

Best Practices for Auto-Batching

1. Start with Default: Use batch_size=25 for most scenarios
2. Adjust for Memory: Decrease batch_size if you still experience issues
3. Combinatorial Mode: Be extra careful - 25 LoRAs × 3 strengths = 75 combinations
4. Save Between Batches: Save your results after each batch to avoid data loss
5. Use Plot Parameters: The batch info appears in plot visualizations for easy tracking

File Naming Patterns

Supported Epoch Formats

• model-v1-000004.safetensors → Epoch 4
• style_epoch_020.safetensors → Epoch 20
• lora-000100.safetensors → Epoch 100

Natural Sorting Examples

Files are sorted intelligently:

1. model-000004.safetensors
2. model-000020.safetensors
3. model-000100.safetensors

Best Practices

Folder Organization

models/loras/
├── my_style/
│   ├── style-000010.safetensors
│   ├── style-000020.safetensors
│   └── style-000030.safetensors
└── character/
    ├── char-v2-000005.safetensors
    └── char-v2-000010.safetensors

Testing Workflows

1. Initial Testing: Use single strength (1.0) to evaluate all epochs
2. Fine-tuning: Use combinatorial mode with multiple strengths
3. Final Selection: Filter to specific epochs and test strength range

Pattern Filtering Examples

# Include only specific versions
include_pattern: "v2|v3"

# Exclude test/backup files
exclude_pattern: "test|backup|old"

# Include specific epoch range
include_pattern: "epoch_0[3-7]0"

Integration with Other Nodes

Common Pipelines

1. LoRA Comparison Grid:

   LoRAFolderBatch → KSampler → Image Grid → Save
   

2. Strength Testing:

   LoRAFolderBatch → PlotParameters → Graph Display
   

3. Combined with FLUX:

   LoRAFolderBatch → FluxSamplerParams → KSampler
   

Tips and Tricks

Memory Management

• Start with fewer LoRAs when testing combinatorial mode
• Use sequential mode for initial epoch evaluation
• Clear LoRA cache between large batch runs

Optimal Strength Ranges

• Style LoRAs: 0.5-1.0
• Character LoRAs: 0.7-1.2
• Detail LoRAs: 0.3-0.7

Debugging

• Check lora_list output to verify correct files were found
• Use lora_count to confirm expected number of LoRAs
• Test patterns with include/exclude before full runs

Troubleshooting

No LoRAs Found

• Verify folder path (relative to models/loras or use absolute)
• Check file extensions (.safetensors)
• Test without filters first

Pattern Not Working

• Patterns use Python regex syntax
• Test patterns in regex tester first
• Case-sensitive by default

Memory Issues

• Reduce batch_count in combinatorial mode
• Process LoRAs in smaller groups
• Use sequential mode for large sets

Advanced Examples

Multi-Version Testing

# Test different versions at different strengths
folder_path: "character_loras"
include_pattern: "v[1-3]"
strength: "0.6, 0.8, 1.0"
batch_mode: combinatorial

Epoch Progression Analysis

# Test every 10th epoch
folder_path: "training_output"
include_pattern: "0[0-9]0\\.safetensors$"
strength: "1.0"
batch_mode: sequential

Version History

• 1.0.0: Initial adaptation from comfyui-essentials-nodes
• 1.0.1: Added natural sorting for epochs
• 1.0.2: Enhanced pattern filtering
• 1.0.3: Improved batch modes and strength parsing
• 1.0.4: Added auto-batching for large LoRA collections

Credits

Original implementation by cubiq in comfyui-essentials-nodes. Adapted and maintained by the ComfyAssets team.