diff --git a/LLavaTagger/README.md b/LLavaTagger/README.md new file mode 100644 index 0000000000000000000000000000000000000000..c9381e430944f9c02cfca1c90acc6158604eb5e9 --- /dev/null +++ b/LLavaTagger/README.md @@ -0,0 +1,21 @@ +# LLavaTagger + +LLavaTagger is a python script that tags images based on a given prompt using the [LLaVA](https://llava-vl.github.io/) multi modal llm. LLavaTagger supports using any number of gpus in ddp parralel for this task. + +## How to use + +first create a python venv and install the required packages into it: + + $ python -m venv venv + $ source venv/bin/activate + $ pip install -r requirements.txt + +Then run LLavaTagger for instance like so: + + $ python LLavaTagger.py --common_description "a image of a cat, " --prompt "describe the cat in 10 to 20 words" --batch 8 --quantize --image_dir ~/cat_images + +By default LLavaTagger will run in parallel on all available gpus, if this is undesriable please use the ROCR_VISIBLE_DEVICES= or CUDA_VISIBLE_DEVICES= environment variable to hide unwanted gpus + +LLavaTagger will then create a meta.jsonl in the image directory sutable to be used by the scripts of [diffusers](https://github.com/huggingface/diffusers) to train stable diffusion (xl) if other formats are desired ../utils contains scripts to transform the metadata into other formats for instace for the use with [kohya](https://github.com/bmaltais/kohya_ss) + +If editing the created tags is desired, [QImageTagger](https://uvos.xyz/git/uvos/QImageTagger) can be used for this purpose diff --git a/PersonDatasetAssembler/README.md b/PersonDatasetAssembler/README.md new file mode 100644 index 0000000000000000000000000000000000000000..e6ec6b81a7682dbc03467a9156cb49331a5602bb --- /dev/null +++ b/PersonDatasetAssembler/README.md @@ -0,0 +1,20 @@ +### PersonDatasetAssembler + +PersonDatasetAssembler is a python script that finds images of a spcific person, specified by a referance image in a directory of images or in a video file. PersonDatasetAssembler supports also raw images. + +## How to use + +first create a python venv and install the required packages into it: + + $ python -m venv venv + $ source venv/bin/activate + $ pip install -r requirements.txt + +Then run PersonDatasetAssembler for instance like so: + + $ python PersonDatasetAssembler.py --referance someperson.jpg --match_model ../Weights/face_recognition_sface_2021dec.onnx --detect_model ../Weights/face_detection_yunet_2023mar.onnx --input ~/Photos --out imagesOfSomePerson + +Or to extract images from a video: + + $ python PersonDatasetAssembler.py --referance someperson.jpg --match_model ../Weights/face_recognition_sface_2021dec.onnx --detect_model ../Weights/face_detection_yunet_2023mar.onnx -i ~/SomeVideo.mkv --out imagesOfSomePerson + diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..9871c1ed5b64df689452f68e5dac2e0cc007aa9d --- /dev/null +++ b/README.md @@ -0,0 +1,35 @@ +# SDImagePreprocess + +This repo contains a collection of high performance tools intended to ease the createion of datasets for image generation AI training like stable diffusion. + +## Included tools + +This repo contains the following tools: + +### SmartCrop + +SmartCrop is an application that uses content aware croping using, [seam carving](https://en.wikipedia.org/wiki/Seam_carving) and resizeing to bring a directory of images into the deisred size and aspect ratio for training. SmartCrop ist configurable to prioritize specific items or specifc persons in the images provided. + +#### Content detected in image: + + + +#### Cropped image based on content: + + +### PersonDatasetAssembler + +PersonDatasetAssembler is a python script that finds images of a spcific person, specified by a referance image in a directory of images or in a video file. PersonDatasetAssembler supports also raw images. + +### LLavaTagger + +LLavaTagger is a python script that tags images based on a given prompt using the [LLaVA](https://llava-vl.github.io/) multi modal llm. LLavaTagger supports using any number of gpus in ddp parralel for this task. + +### DanbooruTagger + +DanbooruTagger is a python script of dubious utility that tags images based using the [DeepDanbooru](https://github.com/KichangKim/DeepDanbooru) convolutional network. + + +## License + +All files in this repo are litcenced GPL V3, see LICENSE diff --git a/SmartCrop/README.md b/SmartCrop/README.md new file mode 100644 index 0000000000000000000000000000000000000000..7a216feceec24f9a0be88317338f975df0b016b6 --- /dev/null +++ b/SmartCrop/README.md @@ -0,0 +1,50 @@ +# SmartCrop + +SmartCrop is an application that uses content aware croping using, [seam carving](https://en.wikipedia.org/wiki/Seam_carving) and resizeing to bring a directory of images into the deisred size and aspect ratio for training. SmartCrop ist configurable to prioritize specific items or specifc persons in the images provided. + +## Requirements + +* [cmake](https://cmake.org/) 3.6 or later +* [opencv](https://opencv.org/) 4.8 or later +* A c++17 capable compiler and standard lib like gcc or llvm/clang +* git is required to get the source + +## Building + +The steps to build this application are: + + $ git clone https://uvos.xyz/git/uvos/SDImagePreprocess.git + $ cd SDImagePreprocess + $ mkdir build + $ cmake .. + $ make + +The binary can then be found in build/SmartCrop and can optionaly be installed with: + + $ sudo make install + +## Basic usage + +To process all images in the directory ~/images and output the images into ~/proceesedImages: + + $ smartcrop --out processedImages ~/images/* + +To also focus on the person in the image ~/person.jpg + + $ smartcrop --out processedImages --focus-person ~/person.jpg ~/images/* + +To also enable seam carving + + $ smartcrop --out processedImages --focus-person ~/person.jpg --seam-carving ~/images/* + +see smartcrop --help for more + +## Example + +#### Content detected in image: + + +#### Cropped image based on content: + + +