ComfyUI Extension: ComfyUI_SimpleButcher

Authored by KLL535

Created 2 months ago

Updated 16 days ago

3 stars

Node to automate batch generation with randomize prompts from text files. It mimics Forge's functionality, allowing you to combine text elements and LoRA. The node supports writing LoRA in any order within a text file using formats like lora:name:1.0 or lora:name:unet=1.0:te=0.75, without needing separate nodes. The node understands LoRA names in Forge's style, when the name is not the filename, but the internal name from the metadata.

Custom Nodes (8)

README

ComfyUI_SimpleButcher

Description:

Node to automate batch generation with randomize prompts from text files. It mimics Forge's functionality, allowing you to combine text elements and LoRA. The node supports writing LoRA in any order within a text file using formats like <lora:name:1.0>, without needing separate nodes. The node understands LoRA names in Forge's` style, when the name is not the filename, but the internal name from the metadata.

Why is this needed:

I needed to automate the process of generating images and could not find any straightforward ready-made solutions for this purpose. Hence, I developed my own solution. The primary idea is to replicate and expand Forge's functionality: specifically, the Prompts from file or textbox script. This will allow you to easily combine pre-prepared lists of prompts, LoRA, styles, allowing you to combine texts as you wish. The resulting images prompt in metadata must be compatible with Forge, ideally appearing no different than those generated by it. However, I encountered a problem because Forge uses the internal name from the LoRA file metadata as the LoRA's name instead of the filename. Consequently, all existing solutions failed to comprehend my templates. I also wish for the ability to write LoRA in any order and quantity within a text file using this format: <lora:name:1.0> or <lora:name:unet=1.0:te=0.75>. This should be automatically applied without the need to create separate nodes for each LoRA. Additionally, I would like it if the CivitAI platform could understand metadata. As the closest solution I found was the alexopus/ComfyUI-Image-Saver project. However, this alternative loses LoRA written in Forge style by using the internal name instead of the filename.

workflow

Caution:

[!] This nodes create a new file in LoRA/Model ditectory:

files *.sha256 next to LoRA, the file contains the hash sum of the LoRA file.
files *.sha256 next to model, the file contains the hash sum of the model file.
file lora_name.json in first LoRA folder (default path ComfyUI\models\loras\).

How to install?

Method 1: Manager (Recommended)

If you have ComfyUI-Manager, you can click Custom Nodes Manager and find ComfyUI_SimpleButcher.

Method 2: Manual

In Windows:

run cmd, go to the ComfyUI folder
cd custom_nodes
git clone https://github.com/KLL535/ComfyUI_SimpleButcher.git
cd ComfyUI_SimpleButcher
.\..\..\..\python_embeded\python.exe -s -m pip install -r requirements.txt
Start/restart ComfyUI

Node 1: Simple Load Line From Text File

Simple tool for loading prompts directly from a text file, enabling automation of the entire batch process. You can combine this nodes to generate fully randomized texts and LoRA, sourced from your own text and LORA templates.

Input:

name - STRING - Just a name. To avoid confusing nodes if there are many of them.
start - INT - Start position for increment or decrement methods.
load_file - BOOLEAN - if True, load center of prompt from file.
file_path - STRING - Path to the file from which the lines for the central part of the prompt will be taken.
next - increment or decrement or fixed or random or random no repetitions - Option for enumerating lines, in case of randomness the start parameter is ignored. In the latter case, the lines will be selected randomly, but without repetitions.
prefix - STRING - Add text to the beginning of the prompt (optional).
postfix - STRING - Add text to the ending of the prompt (optional).

Output:

text - STRING - Output text based on the principle: prefix + line from file + postfix.
batch_counter - INT - Current batch counter, if you start a new batch the counter will reset to 1.
line_counter - INT - Current line counter from file.
lines - INT - Total lines in file. If batch is greater than lines in file, they will be read in a loop.

Node 2: Simple Extract Lora From Text

If the input text includes LoRA written in Forge style, such as <lora:name:1.0> or <lora:name:unet=1.0:te=0.75>, this node automatically segregates the text into separate prompts and LoRA.

Input:

text - STRING - Text with mixed lora and prompt.

Output:

lora_text - STRING - The text contains only lora
prompt - STRING - The text contains only prompt

Node 3: Simple Lora Loader

This tool can load multiple LoRAs described in Forge style, such as <lora:name:1.0> or <lora:name:unet=1.0:te=0.75>. It recognizes both file names and internal LoRA names obtained from the metadata of *.safetensors files via the ss_output_name field. When the workflow is first run, a dictionary lora_name.json is generated to store the lora paths, internal names and hash, and is saved in the LoRA folder to speed up future operations. Hash values for all LoRA are calculated once and stored in *.sha256 files next to each LoRA. If there are many files and the hash has not been calculated before, this process may take a long time. If you update LoRAs place, add new LoRA, simply click button Update LoRA dictionary and run workflow, this will update the file lora_name.json.

Input:

model - MODEL - The diffusion model the LoRA will be applied to.
clip - CLIP - The CLIP model the LoRA will be applied to.
lora_text - STRING - Multiple LoRAs discriptions in Forge style: <lora:name:1.0> or <lora:name:unet=1.0:te=0.75>
multiple_strength_unet - FLOAT - Multiple of unet strength. If there is a lot of lore, they can spoil the model, it is possible to multiply the weight of all incoming lore by this coefficient.
multiple_strength_clip - FLOAT - Multiple of clip strength.
limit_strength_unet - FLOAT - Max of unet strength. If there are a lot of lore, they can spoil the model, it is possible to limit the weight of all incoming lore to this number.
limit_strength_clip - FLOAT - Max of clip strength.
Update LoRA dictionary - Button - Click this button and run workflow will update the file lora_name.json (if you update LoRAs place or add new LoRA).

Output:

model - MODEL - Output diffusion model.
clip - CLIP - Output CLIP model.
civitai_lora - STRING - List of loras and their weights that were applied to the model.
civitai_lora_hash - STRING - The first 10 characters of the hash sum of the loras, needed for the CivitAI site.

Node 4: Simple Image Saver (as Forge)

Image Saver designed for saved metadata in Forge-style, allowing retrieval of LoRAs' names and hashes directly from the Simple LoRA Loader node. This tool automatically organizes output images into subfolders based on the date. The filenames of the saved files include a sequence number and seed. Hash values for current model are calculated once and stored in *.sha256 files next to model.

Input:

images - IMAGE - image(s) to save.
output_path - STRING - Path where images will be saved.

Optional metadata (as Forge).

If not connected these lines will be missing:

prompt_text - STRING.
SEED - INT. If the override_parameters input is connected, the seed will be read from there.
modelname - STRING - Use Checkpoint Loader with Name (Image Saver) node.
steps - INT.
sampler - any - use Sampler Selector (Image Saver) node.
schedule - any - use Scheduler Selector (Image Saver) node.
CFG_scale - FLOAT.
distilled_CFG_scal - FLOAT.
width - INT.
height - INT.
beta_schedule_alph - FLOAT.
beta_schedule_beta - FLOAT.
civitai_lora - STRING - Loras name from Simple Lora Loader node.
civitai_lora_hash - STRING - Loras hash from Simple Lora Loader node.
negative - STRING - negative prompt if needed.
override_parameters - STRING - replaces all above parameters, if they have already been collected.
override_workflow - STRING - replaces the current workflow. Json text.
override_prompt - STRING - replaces the current prompt data. Json text.
save_comfy_workflow - BOOLEAN - if True, saves the current workflow to metadata.
save_comfy_prompt - BOOLEAN - if True, saves the current prompt data (not prompt_text) to metadata.

Output:

metadata_text - STRING - Metadata written to file.

Node 5: Simple Load Image With Metadata

This node is made to upload images with all metadata, including Forge's metadata, so as not to lose them. As far as I understand, there are three metadata fields:

parameters - A convenient representation. It is easy to open the image metadata and read the prompt.
workflow - The workflow is stored here, so that later it can be loaded by dragging the file, in principle, it is convenient. But it is very difficult for a person to read this.
prompt - I do not understand what is stored here. It is very difficult for a person to read this.

Output:

image - IMAGE - load image
mask - MASK - mask of image
metadata_parameters (forge) - STRING - metadata Forge.
metadata_workflow - STRING - workflow Comfy-ui. Json text.
metadata_prompt - STRING - metadata Comfy-ui. Json text.

A simple way to re-save an image without losing metadata. If you connect a workflow, you can save the previous workflow. If you do not connect a workflow, the current workflow will be saved.

Node 6: Simple Load Images from Dir

Same as for text, only for batch images from a directory.

Advantages over other solutions:

Supports non-repetition random mode (to go through all images without repetitions, but randomly, the batch count should be equal than the number of images).
Supports output of raw metadata (including Forge metadata) to avoid losing it.

Input:

name - STRING - Just a name. To avoid confusing nodes if there are many of them.
input_path - STRING - Path to the dir to images.
start - INT - Start position for increment or decrement methods.
next - increment or decrement or fixed or random or random no repetitions - Option for enumerating files, in case of randomness the start parameter is ignored. In the latter case, the files will be selected randomly, but without repetitions.
include_subdir - BOOLEAN - Turn on to read all images and subdirectories too.

Output:

image - IMAGE - load image
mask - MASK - mask of image
metadata_parameters (forge) - STRING - metadata Forge.
metadata_workflow - STRING - workflow Comfy-ui. Json text.
metadata_prompt - STRING - metadata Comfy-ui. Json text.
batch_counter - INT - Current batch counter, if you start a new batch the counter will reset to 1.
line_counter - INT - Current file counter.
lines - INT - Total files. If batch is greater than files, they will be open in a loop.

Node 7: Simple Auto Bypass

Sometimes when I disable a whole group, the information coming into it passing through many complex nodes is lost (becomes None), and I need this information to pass through the disabled group without changes. Here is such a simple solution. I later found an alternative: Any Switch (rgthree). But there are 6 input groups, and I have two - that's the only difference.