MFLUX v.0.6.0 Release Notes

Major New Features

🌐 Third-Party HuggingFace Model Support

• Comprehensive ModelConfig refactor to support compatible HuggingFace dev/schnell models
• Added ability to use models like Freepik/flux.1-lite-8B-alpha and shuttleai/shuttle-3-diffusion
• New --base-model parameter to specify which base architecture (dev or schnell) a third-party model is derived from
• Maintains backward compatibility while opening up the ecosystem to community-created models

🎭 In-Context LoRA

• Added support for In-Context LoRA, a powerful technique that allows you to generate images in a specific style based on a reference image without requiring model fine-tuning
• Introduced a new command-line tool: mflux-generate-in-context
• Includes 10 pre-defined styles from the Hugging Face ali-vilab/In-Context-LoRA repository
• Detailed documentation on how to use this feature effectively with prompting tips and best practices

🔌 Automatic LoRA Downloads

• Added ability to automatically download LoRAs from Hugging Face when specified by repository ID
• Simplifies workflow by eliminating the need to manually download LoRA files before use

🧠 Memory Optimizations

• Added --low-ram option to reduce GPU memory usage by constraining the MLX cache size and releasing text encoders and transformer components after use
• Implemented memory saver for ControlNet to reduce RAM requirements
• General memory usage optimizations throughout the codebase

🗜️ Enhanced Quantization Options

• Added support for 3-bit and 6-bit quantization (requires mlx > v0.21.0)
• Expanded quantization options now include 3, 4, 6, and 8-bit precision

⚠️Breaking changes

Previously saved quantized models will not work for v.0.6.0 and later. See #149 for more details.

Interface Improvements

🔧 Modified Parameters

• The previous --init-image-path parameter is now --image-path
• The previous --init-image-strength parameter is now --image-strength

🖼️ Image Generation Enhancements

• Added --auto-seeds option to generate multiple images with random seeds in a single command
• Added option to override previously saved test images
• Added --controlnet-save-canny option to save the Canny edge detection reference image used by ControlNet
• Improved handling of edge cases for img2img generation

🔄 Callback System

• Implemented a general callback mechanism for more flexible image generation pipelines
• Added support for before-loop callbacks to accept latents
• Enhanced StepwiseHandler to include initial latent

Architecture Improvements

🏗️ Code Refactoring

• Removed 'init' prefix for a more general interface
• Removed ConfigControlnet - the controlnet_strength attribute is now on Config
• Simplified quantization by removing unnecessary class predicates
• Refactored model configuration system
• Refactored transformer blocks for better maintainability
• Unified attention mechanism in single and joint attention blocks
• Added support for variable numbers of transformer blocks
• Optimized with fast SDPA (Scaled Dot-Product Attention)
• Added PromptCache for small optimization when generating with repeated prompts

🧰 Developer Tools

• Added Batch Image Renamer tool as an isolated uv run script
• Added descriptive comments for attention computations

Compatibility Updates

• Updated to support the latest mlx version
• Fixed compatibility issues with HuggingFace dev/schnell models

Bug Fixes

• Fixed handling of edge cases for img2img generation
• Various small fixes and improvements throughout the codebase

Contributors

• @anthonywu
• @ssakar
• @azrahello
• @DanaCase