Hello! This section is devoted to the following problem: large language models (LLMs) are unable to perform highly specialized image classification and generation tasks. One such domain is armwrestling techniques. Accordingly, we attempt to address this problem.
First, the following armwrestling techniques were selected in accordance with the armwrestling triangle: toproll, hook, press, and kingsmove. The datasets directory contains: 640 images of armwrestling techniques (160 per technique), 80 validation images of armwrestling techniques (20 per class), as well as 400 images containing techniques (50 per arm). The technique is defined for the person on the left.
Second, the classification directory contains programs for training and inference of a model that classifies images with armwrestling techniques. For training, train_dir contains training images, val_dir contains validation images, and out_dir specifies the path for saving model weights. For inference, the path to the weights and the path to the images must be specified. Image filenames must follow the format (example): toproll_001.jpg (or hook, press, kingsmove). A comparison of classification quality between gpt-5.2-pro and the trained model is shown below (GPT on the left, trained model on the right).
Снимок экрана 2026-01-04 193821
Third, SDXL was fine-tuned using LoRA. The input_root_dir contains subfolders (for example, left_hook) with images, work_dir is the directory where the weights are saved, and folder_from and folder_to specify which folders are used for training (separate weights are trained for each folder). For inference, the lora_path must be specified. The steps and guidance_scale parameters can be adjusted. The results are shown below.
collage_arm
For classification use square-shaped images. If you have any questions, please contact aranibargeorge@gmail.com Here you can read my article (...soon...)
