FLUX-LoRA-DLC

Enables fine-tuning of FLUX text-to-image model weights through Low-Rank Adaptation (LoRA), a parameter-efficient training technique that freezes base model weights and trains only low-rank decomposition matrices. The implementation uses gradient-based optimization on image generation tasks, allowing users to customize model behavior for specific visual styles, subjects, or artistic directions without full model retraining. Training state is managed through HuggingFace Spaces infrastructure with Gradio UI for parameter configuration.

Provides a Gradio-based UI running on HuggingFace Spaces that exposes LoRA training parameters (rank, learning rate, steps, batch size) and generates preview images at configurable intervals during training. The interface handles file uploads for training datasets, manages training job lifecycle (start/pause/resume), and displays loss curves or training metrics in real-time. State is persisted in the Spaces environment with outputs downloadable as .safetensors files.

Manages trained LoRA adapter export in .safetensors format with embedded metadata (training config, model version, LoRA rank/alpha values). The system ensures compatibility by storing model architecture information and version tags, allowing exported weights to be loaded into compatible FLUX inference pipelines. Export includes optional quantization or compression options to reduce file size for distribution.

Provides utilities to preprocess uploaded image datasets for LoRA training, including resizing to FLUX-compatible dimensions (typically 768x768 or 1024x1024), format conversion (PNG/JPG to standardized format), and optional augmentation (random crops, flips, color jitter). The system validates image quality, filters corrupted files, and generates captions or prompts for each image using vision-language models or user-provided text. Augmentation parameters are configurable to control dataset diversity without manual image editing.

Tracks training metrics (loss, learning rate schedule, gradient norms) during LoRA training and visualizes them in real-time through interactive plots (loss curves, learning rate decay, validation metrics if applicable). The system logs training events to a structured format (JSON or CSV) for post-training analysis and reproducibility. Metrics are displayed in the Gradio interface with configurable refresh intervals, and historical training runs can be compared side-by-side.

Loads trained LoRA weights and applies them to the base FLUX model for image generation, merging low-rank adapter matrices with frozen base weights during inference. The system supports prompt-based generation with optional negative prompts, seed control for reproducibility, and guidance scale adjustment for prompt adherence. LoRA inference is implemented as a forward pass modification that adds adapter outputs to base model activations, with minimal latency overhead compared to base model inference.