LyCORIS Chopping #

LoRAs can sometimes include unwanted elements or behaviors during training. “Chopping” allows you to selectively enable or disable different parts of a LoRA model to fine-tune its effects. This can help control style transfer, character consistency, and other attributes.

Quick Solution: Block Weighting #

You can use block weighting tools during generation:

• ComfyUI Inspire Pack - Includes LoRA block weight functionality
• A1111 LoRA Block Weight

Permanent Solution: Chopping #

For a permanent solution, you can use chop_blocks.py by Gaeros to modify the LoRA file itself:

git clone https://github.com/elias-gaeros/resize_lora
cd resize_lora

How to Use #

python chop_blocks.py --model input.safetensors --save_to output.safetensors --vector "1,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1"

Understanding Block Weights #

Block weight vectors control which layers are kept or removed using values from 0 to 1:

• 1.0 = Keep layer completely
• 0.0 = Remove layer completely
• Values between = Partial effect

SDXL Block Structure (27 Values) #

SDXL models use a 27-element vector with this mapping:

• Indices 0-8: UNet Down Blocks
• Indices 9-11: UNet Mid Block
• Indices 12-20: UNet Up Blocks
• Indices 21-26: CLIP Text Encoder

By setting values to 1, you enable training for that block; setting to 0 disables it.

Common Block Weight Presets #

My Personal Block Weight Presets #

LyCORIS Preset/Config System #

The Preset/Config system was added after LyCORIS 1.9.0 for more fine-grained control during training.

Built-in Presets #

LyCORIS provides several presets that can be applied with --network_args "preset=NAME":

• preset=full: Default preset, trains all layers in the UNet and CLIP
• preset=full-lin: Same as full but skips convolutional layers
• preset=attn-mlp: The “kohya preset”, trains all transformer blocks
• preset=attn-only: Only attention layers will be trained (common in research papers)
• preset=unet-transformer-only: Same as kohya_ss/sd_scripts with disabled TE, or attn-mlp preset with train_unet_only enabled
• preset=unet-convblock-only: Only ResBlock, UpSample, and DownSample will be trained

Custom Configs with TOML #

For more detailed control, you can create a config.toml file and specify module targets and algorithms. This allows you to:

• Choose different algorithms for specific module types/modules
• Use different settings for specific module types/modules
• Enable training for specific module types/modules

The config system supports the following arguments:

• enable_conv (bool): Enable training for convolution layers or not
• unet_target_module (list[str]): Classes of modules to train in UNet
• unet_target_name (list[str]): Names of modules to train in UNet (regex supported)
• text_encoder_target_module (list[str]): Classes of modules to train in Text Encoder
• text_encoder_target_name (list[str]): Names of modules to train in Text Encoder (regex supported)
• module_algo_map/name_algo_map (dict[str, str]): Apply different settings to different class/name

Implementing Block Weights #

There are several ways to apply block weights to your LyCORIS models.

During Training #

When training with sd-scripts, you can specify block weight vectors using the --block_weights argument:

accelerate launch train_network.py \
  --network_module=lycoris.kohya \
  --network_args "algo=locon" \
  --block_weights="0,0,1,1,1,0,0,0,0,1,1,1,0,0,0,0,1,1,1,0,0,1,1,1,1,1,1" \
  [other training arguments]

Alternatively, you can apply presets during training using the --network_args parameter:

accelerate launch train_network.py \
  --network_module=lycoris.kohya \
  --network_args "algo=locon" "preset=attn-mlp" \
  [other training arguments]

During Generation #

In ComfyUI #

In ComfyUI, you can use the LoRA Loader (Block Weight) node to input block vectors directly.

In Automatic1111 (A1111) #

In A1111, you can incorporate block weight vectors into your prompts using this syntax:

<lyco:"model_name":1:1:lbw=0,0,1,1,1,0,0,0,0,1,1,1,0,0,0,0,1,1,1,0,0,1,1,1,1,1,1>

Replace “model_name” with the actual name of your LyCORIS model.