Efficient Loader Node Documentation

Overview

The Efficient Loader node is a custom node for ComfyUI designed to streamline workflows by loading and caching various types of models. It supports the loading of Checkpoint, VAE, and LoRA models and provides additional features to enhance the overall workflow efficiency.

Features

• Load & Cache Models: The node can load and cache Checkpoint, VAE, and LoRA type models, which helps in reducing the load times and improves workflow efficiency.
• Prompt Encoding Options: The node includes options for positive and negative prompt text boxes. It provides encoding options to control how prompts are processed using different token normalization and weight interpretation methods.
• Compatibility: The Efficient Loader node is used by the XY Plot node for many of its plot type dependencies.
• Custom Menu Options: The node comes with a variety of custom menu options, enhancing user control over the modeling process.

Inputs

The Efficient Loader node accepts the following inputs:

1. Checkpoint Name (ckpt_name): Specifies the checkpoint model to load.
2. VAE Name (vae_name): Determines which VAE model to use; includes an option for baked VAE.
3. Clip Skip (clip_skip): An integer input that adjusts the layers read from a Clip model.
4. LoRA Name (lora_name): Indicates which LoRA model to load; "None" if no LoRA is used.
5. LoRA Model Strength (lora_model_strength): Determines the model strength for the LoRA.
6. LoRA Clip Strength (lora_clip_strength): Sets the clip strength for the LoRA.
7. Positive Prompt (positive): String input specifying the positive conditioning prompt.
8. Negative Prompt (negative): String input specifying the negative conditioning prompt.
9. Token Normalization (token_normalization): Specifies the method to normalize tokens.
10. Weight Interpretation (weight_interpretation): Method to interpret weights within the prompt.
11. Empty Latent Width (empty_latent_width): Width for creating an empty latent tensor.
12. Empty Latent Height (empty_latent_height): Height for creating an empty latent tensor.
13. Batch Size (batch_size): Sets the number of samples to generate in a batch.
14. LORA Stack (lora_stack) - Optional: Stack of LoRA models.
15. Control Net Stack (cnet_stack) - Optional: Stack of ControlNet models.

Outputs

The Efficient Loader node produces the following outputs:

1. Model: The loaded model object.
2. Positive Conditioning: The positive text encoding and conditioning data for the model.
3. Negative Conditioning: The negative text encoding and conditioning data for the model.
4. Latent: A latent tensor initialized with zeros to facilitate model generation.
5. VAE: The VAE model that was loaded or baked.
6. Clip: The Clip model associated with the Checkpoint and LoRA models.
7. Dependencies: Metadata for plot type dependencies, useful for the XY Plot node.

Use in ComfyUI Workflows

In ComfyUI workflows, the Efficient Loader node is integral for preparing models and conditioning data. It is especially valuable for workflows involving multiple models, such as Checkpoints and LoRAs, as it efficiently manages model resources by caching them. This ensures faster computations by minimizing repeated loading of models. Its ability to transform and cache inputs makes it essential for workflows that involve complex iterative processes like model tuning, image synthesis, or any task requiring varied model inputs.

Special Features

• Custom Menu Options: The Efficient Loader features a variety of custom menu options that enhance functionality and user experience in ComfyUI.

• Prompt Control: The node not only allows for normal prompt encoding but also lets users specify methods for token normalization and weight interpretation to refine how prompts influence model behavior.

• Model Stacking Compatibility: It supports stacking inputs from multiple LoRA and ControlNet nodes, enabling advanced configurations and refined control.

Overall, the Efficient Loader node is a versatile tool that enhances the capability of ComfyUI workflows by optimizing loading processes, providing encoding options, and supporting a wide array of model configuration setups.