This repository contains custom nodes for ComfyUI.
Example workflow in the prompt_with_styles.json file and mikey_node_workflow.json
This node allows you to prompt with a style and loras. You can add a lora with <lora:lora_name> or <lora:lora_name:0.8> syntax for a weight of 0.8 as an example.
You can add styles by using the <style:style_name> syntax.
Some of the styles built in are:
SAI-Enhance, SAI-Anime, SAI-Photographic, SAI-Digital art, SAI-Comic book, SAI-Fantasy art, SAI-Analog film, SAI-Neonpunk, SAI-Isometric, SAI-Lowpoly, SAI-Origami, SAI-Line art, SAI-Craft clay, SAI-Cinematic, SAI-3d-model, SAI-pixel art, SAI-Texture, photographic, deep-field, analog-film, cinematic, red, documentary, nature-photography, editorial, polaroid, 3d-model, low-poly, modeling-compound, diorama, miniatures, abstract, anime, comic-book, digital-art, fantasy-art, dark-fantasy, oil-painting, watercolor, charcoal, color-pencil, crayon, cross-stitch, felt, origami, scrapbook, isometric, line-art, vector-art, neon-punk, pixel-art, tile-texture, lithography, mosaic, woodblock-print, sticker, stained-glass, tattoo, statue, album-art, wes-anderson, vaporwave, clarendon, gingham, juno, lark, nonagon, kaleidoscope, haunting, glam, mecha, padam, collage, paper-mache, macrame, quilling, pottery, batik, felting, marquetry, wildstyle-graffiti, bubble-graffiti, stencil-graffiti, throw-up-graffiti, tag-graffiti
There is a V1 and V2 Prompt with Style node with different functionality, style selection instead of style syntax, no clip conditioning. Check them out if you like to break your workflow up into more nodes.
Wildcards are supported using the __word__ syntax. Folder location is comfyui/wildcards You can select more than 1 line from the wildcard by using the syntax 2$$__wildcard__ for 2 lines in this example. You can also add a word to search for in the wildcard file e.g. __wildcard|word__
You can also make sure it selects the same line from the wildcard by using the __!wildcard__ syntax, not the ! before the filename. You can also use __+wildcard__ __-wildcard__ and __*wildcard__ to use the offset + 1 or - 1, or randomly select a different line.
You can add your own styles to a file named user_styles.json in the root ComfyUI folder. You have to create this file yourself.
The format needs to look like the following example:
{
"styles": {
"newspaper": {
"positive": "newspaper comic strip, panels, black and white",
"negative": "photograph, color"
}
}
}You can also add your own custom ratios to a file named user_ratios.json in the root ComfyUI folder. You have to create this file yourself.
The format needs to look like the following example:
{
"ratios": {
"test": {"width": 1024, "height": 1024}
}
}
Example workflow: mikey_node_workflow.json
This is a sampler node as a companion to the Prompt with Style node, to allow for very straightforward image generation with SDXL.
The node uses a base -> refiner -> upscale with model -> base to produce the final image. It has an inbuilt image complexity detection function to make sure there aren't too many steps run on the final base sampler which can scramble faces and other simple areas on a large image.
The Base Only version of the sampler uses a slightly different approach. It uses a base -> base -> upscale with model -> base to produce the final image. The other key difference with the Base Only version is it skips a step in the middle and adds it to the end. This is a trick which adds extra detail to the image. The smooth step number will add even more extra steps to the end. This might be required if you have an image with a plain background and are seeing some spotty effects created by the skip step method. Usually only 1 extra step is required to clean up the spots. You can also go -1 for very busy images for even more detail.
model_name is for the image uspcale model. These can be found at OpenModelDB and place them in the ComfyUI\models\upscale_models directory. I recommend lollipop as it provides a good balanced image that is not too sharp. If you want lots of sharp details, then try the NMKD Superscale or 4x-UltraSharp.
The seed is used by the samplers.
The upscale_by setting is how large you want the final image to be before it runs the last base sampler over the image. A value of 1 will output the original image size. 2 will be 2x the size of the original.
The highres_strength setting allows you to control this last stage and how much detail it adds. Setting of 1 is the sweet spot I found didn't scramble faces in a few test images. Dial it down if you are still getting scrambled images, or turn it up if you want to add more detail.
These samplers use a tiled approach to resample the image after the upscale model has upscaled the image, to help enhance the image with more details.
This chonky node is for injecting the style prompt into your conditioner by averaging the conditioner passed through the node.
A strength value of 0.5 would be even split between the conditioner passed in and the style selected.
This node allows you to select from a range of different aspect ratios.
This node allows you to enter your own aspect ratio or image size from wherever, and it will make it fit under 1024x1024 resolution.
This node will read an optional user_ratio_presets.json file in the root directory of ComfyUI. This is an example of a preset:
{
"ratio_presets": {
"1024x1024 (AR 1:1 / DEC 1.0:1)": {
"custom_latent_w": 1024,
"custom_latent_h": 1024,
"cte_w": 1024,
"cte_h": 1024,
"target_w": 4096,
"target_h": 4096,
"crop_w": 0,
"crop_h": 0
},
}
There is a file created by masslevel included in mikey_nodes directory called ratio_presets.json. It contains all of the recommended resolutions from SAI and the correct clipt text encoder and target sizes. This file will automatically be read and populated into the selection widget.
It is a really powerful and convenient ratio selector, because you can create all sorts of combinations for the latent image size, clip text encoder image size, target size and crop coordinates.
The swap axis option will rotate the dimensions by 90 degrees if set to true.
If use preset seed is set to true, it will use the seed number to select one of the available presets. This can be used to cycle or randomly select presets, which is super convenient.
This node is the big brother to Preset Ratio Selector. It let's you select a preset like with Preset Ratio Selector, but there is also lots of options for customisation.
When creating a latent image you need to set the width and height. But the Clip Text Encoder for SDXL Base Model also requires width and height, target width and height, and crop width and height.
This node has options to setup the dimensions for each of these using different methods:
- Select a ratio from the select widget, or choose custom to enter your own dimensions.
- If you select custom, you need to fill in the width and height options.
- If you enter a number into the mult input e.g. cte_mult (clip text encoder) it will multiply the latent width and height to create the width and height for that dimension.
- If you enter a number into the res input, it will create a width and height that fits into the res^2 image size while maintining the same ratio as the latent width and height.
- If you enter a number into the fit_size input, it will create a width and height that has the largest dimension equal to fit_size while maintaining the same ratio as the latent width and height.
This node allows you to resize an image to fit into 1024x1024 resolution. Good for Img2Img workflows.
Given a path to a folder containing images, it will resize all images in the folder to fit the 1024^2 resolution and feed into the workflow. Careful of folders with lots of images!
This node allows you to save an image with the prompt data in the filename.
The filename will start with a datestamp, then part of the positive prompt.
It will also save the positive prompt and negative prompt to the png data.
Made for the Ultimate SD Upscaler to help calcuate tile size to evenly fit inside the larger image and not get weird edges.
The resolution is what is suitable for the model you are using, 1.5: 512, 2.1: 768, SDXL Base: 1024, SDXL Refiner: 768
This node is for building a string using your wildcard files.
I want to give a shout-out to masslevel on discord for the great idea to have nodes for metadata.
This node lets you add more metadata to the image that is passed through this node in your workflow. The metadata should be written to the image file in the Save Image node.
The label is the key or label you want to give to the metadata, and the text box is for entering a value. Convert these to inputs if you want to pipe in string values from other nodes.
This node should pick up the info stored in the prompt and extra_pnginfo data and save it to a txt file
This will apply a HaldCLUT to an image to change the colors, which tend to imitate the look of the film or filter. I have included some in this package, but you can find more png files at rawtherapee.com
This node will add a caption bar to the bottom of the image. Useful for adding a prompt or the name of the checkpoint to the image.
If you create a font folder in the base directory of comfyui e.g. C:\ComfyUI\fonts and place ttf files in that folder, the widget will automatically be a drop down selection with the files in that directory. Otherwise it will be a text input for you to put the full path to the font file.
This node creates a random number or Seed. It outputs the seed as a Integer and also a String version which can be used to add to metadata or wherever you want.
Converts an integer to a string.
Converts a float to a string.
- VAE Decode 6GB
To use these nodes, simply open a terminal in ComfyUI/custom_nodes/ and run:
git clone https://github.com/bash-j/mikey_nodes
This project is licensed under the MIT License - see the LICENSE file for details.