LoRa Dataset Tools

A custom node for ComfyUI providing nodes to facilitate the process of image gathering,filtering and processing when creating a training dataset to use for training loras.

Nodes

Directory Loader

Loads all images from a directory in a batch torch tensor. The width and height of all images are upscaled to match those of the largest image file in the directory using PIL's Image.LANCZOS

Directory Selector

Loads all images from a direcotry as list of torch tensors, allowing the user to preview images like the built-in "Preview Image" node and select one from the list to be the node's output.

Filter Images Without Faces

Takes as input a batch of images and returns only those which contain at least one face in it using facenet_pytorch's implementation of MTCNN

Load Gallery Manager

Custom node which loads all images from a directory to a Gallery Manager object which can perform some custom functions defined in the extension. Warning: The operations performed with Gallery Managers take place in the OS file system using the charloratools package, for more information please check the package's github. While using the operation nodes in ComfyUI you can expect their return inside the "output" folder. Uses the Preview Image Node to show the images that were batched in the UI.

Gallery Manager to Batch & Batch to Gallery Manager

Custom nodes which allow easy conversion between GalleryManager outputs and torch's tensor batches of images

Add Gallery Manager

Custom node which adds all images from 2 Gallery Manager nodes into a single torch image batch tensor (Will duplicate the repeated images in the result.), Also Returns a Gallery Manager Object.

Filter Images Without Specific Face

Takes as input a batch of images and a single reference image and returns images from the batch which contain the reference image using facenet_pytorch's implementation of MTCNN and InceptionResnetV1

Detect Faces and Draw Detection Boxes

Takes as input a batch of images and returns only those which contain at least one face in it using facenet_pytorch's implementation of MTCNN. The returned images have the MTCNN's predicted detection boxes drawn using opencv-python library.

Delete Duplicates

Compares the hashes of all images in a batch and deletes the images with repeated hashes. Takes as input the torch tensor image batch and a string representing which hashing function to use. The available ones are : sha256, P-Hash,D-Hash,Average-Hash and Crop-Resistant-Hash

Save As HTML Image Gallery

Generates a standalone HTML image gallery from a torch tensor batch of images. The gallery is saved in the output's folder as generated_img_gallery.html or generated_img_gallery_{unique_datetime_string} if there's already a file with that name. The gallery uses js and css distributed by public cdn's to improve the display experience.

Filter Images With Multiple Faces

Filters images with multiple faces - Takes as input a torch image batch and returns only those which have only a single face on it.

Roadmap

• Implement the rest of charloratools functionalities.