EasyLLLite(LLLiteLoader)

The EasyLLLite node (LLLiteLoader) loads an LLLite model and applies it to a base diffusion MODEL, using a conditioning image to inject the LLLite behavior during sampling. It returns the patched MODEL, ready to pass to a sampler node. The node is listed under EasyUse/Loaders.

Description

LLLite is a lightweight attention-patching technique related to LoRA. Instead of training a full LoRA, it uses a conditioning image to apply a visual condition to the model's attention layers. This loader performs the patch before sampling, so no special sampler settings are required.

Inputs

model

The base MODEL that will be patched with the LLLite weights. This is usually the output of a checkpoint loader or another model loader.

model_name

Selects the LLLite model file from available model files. This is a dropdown list populated from the LLLite model folder.

cond_image

The conditioning image used by the LLLite patch. This image provides the visual structural guidance that the LLLite model was trained to condition on.

strength

Controls how strongly the LLLite patch affects the model.

• Default: 1
• Range: 0 to 10
• Step: 0.01

Use 0 to disable the effect, lower values for a subtler influence, and higher values for a stronger influence.

steps

Limits the LLLite patch to a fixed number of sampling steps.

• Default: 0
• Range: 0 to 200

0 means no fixed step-count limit. A positive value restricts the patch to the first N steps of sampling.

start_percent

Sets the starting point of the sampling schedule where the LLLite patch becomes active.

• Default: 0
• Range: 0 to 100
• Step: 0.1

The value is a percentage of the full sampling process. 0 means the patch is active from the beginning.

end_percent

Sets the ending point of the sampling schedule where the LLLite patch stops being active.

• Default: 0
• Range: 0 to 100
• Step: 0.1

With the default 0, the patch remains active until the end of sampling. To focus the patch on part of the generation, set this to a value below 100, for example 30 to stop the patch early.

Outputs

MODEL output

The patched model. Use this output wherever you would normally use the original model input, such as in a KSampler or another model-consuming node.

Usage Notes

• Set strength to 0 if you want to keep the node in the graph but temporarily disable the patch.
• Use steps, start_percent, and end_percent to control when the patch is applied. Leaving them at their defaults keeps the patch active for the entire sampling process.
• The start_percent and end_percent values are percentages, not raw step numbers, so they automatically adapt to samplers with different total step counts.
• If steps is set to a positive value, it restricts the active window to that many sampling steps. The percentage bounds provide an additional way to define the active window.