Diffusion Fine-tuning Optimization Tool Bracket

Bracket is an open-source hyperparameter search tool that automates the repetitive manual experimentation process when fine-tuning Diffusion models (SDXL, Flux.1, SD3.5, HunyuanVideo, etc.).

Key Features and Characteristics:

• Automated Search and Dual Scoring: Uses Optuna TPE to search parameters, combining training loss trends with VLM (Vision Language Model)-based image quality evaluation (prompt adherence, visual quality, etc.) to determine the optimal settings.
• Statistical Confidence: After experiments conclude, it applies a Welch's t-test to provide a Markdown report of settings with a statistically significant advantage, rather than relying on mere subjective judgment.
• Efficient Resource Management: Automatically detects the GPU's VRAM capacity to suggest only feasible settings, and uses a Warm-start feature that runs validated settings first, generating meaningful comparison data from the early stages of the search.
• High Compatibility: Since it runs musubi-tuner and sd-scripts as subprocesses, it can use the existing Kohya training environment's arguments (args) and outputs as-is.