Add Stable Diffusion lora pass

[xiaoyu-work]

This pull request introduces significant improvements to the SD-LoRA data preprocessing pipeline, particularly enhancing support for DreamBooth-style datasets and improving the modularity and usability of related components. The main changes include adding explicit DreamBooth class image handling to both aspect_ratio_bucketing and image_resizing, refactoring dataset loading for HuggingFace datasets, and registering the SDLoRA pass in the configuration.

Enhancements to SD-LoRA data preprocessing and configuration:

DreamBooth support and preprocessing improvements:

• Added explicit processing of DreamBooth class images in both aspect_ratio_bucketing and image_resizing, including resizing, bucket assignment, and output path management. This ensures that class images are handled consistently with instance images and that relevant metadata is tracked for downstream tasks. [1] [2]
• Refactored the image_resizing function to use a helper for per-image processing, improved crop coordinate calculation, and ensured bucket assignment metadata is stored for all processed images. [1] [2] [3] [4]

Codebase and API improvements:

• Updated the olive.data.component.sd_lora.__init__.py to explicitly export key preprocessing modules, making them more discoverable and easier to import elsewhere in the codebase.
• Refactored ImageDataContainer to extract HuggingFace-specific parameters earlier in the dataset loading process, ensuring proper conversion and avoiding parameter leakage to downstream loaders.

Configuration updates:

• Registered the SDLoRA pass in olive_config.json, specifying its dependencies, supported providers, and dataset requirements, enabling its use in Olive pipelines.

[Copilot]

Pull request overview

This PR introduces a comprehensive SD-LoRA training pass that enables fine-tuning of Stable Diffusion models (SD 1.5, SDXL, and Flux) using LoRA adapters, with full support for DreamBooth-style training with prior preservation.

Key changes:

• New SDLoRA pass with automatic model type detection and support for three diffusion architectures (SD 1.5, SDXL, Flux)
• Enhanced data preprocessing pipeline with explicit DreamBooth class image handling in both aspect ratio bucketing and image resizing components
• Validation improvements to DiffusersModelHandler with explicit model checking via model_index.json

Reviewed changes

Copilot reviewed 10 out of 12 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
olive/passes/diffusers/lora.py	New SDLoRA pass implementation with training loops for SD/SDXL (UNet-based) and Flux (DiT-based) models, including prior preservation loss for DreamBooth
olive/data/component/sd_lora/aspect_ratio_bucketing.py	Added explicit processing of DreamBooth class images including bucket assignment, resizing, and metadata tracking
olive/data/component/sd_lora/image_resizing.py	Refactored to support class image processing with helper functions for per-image operations and crop coordinate calculation
olive/data/component/sd_lora/init.py	Explicit exports of preprocessing modules for improved discoverability
olive/data/container/image_data_container.py	Refactored HuggingFace dataset loading to extract ImageDataContainer-specific parameters before passing to base loader
olive/model/handler/diffusers.py	Added validation via is_valid_model() checking for model_index.json, plus adapter_path property for LoRA weights
olive/olive_config.json	Registered SDLoRA pass with GPU-only support and sd-lora extra dependencies
test/passes/diffusers/test_lora.py	Comprehensive test suite covering SD 1.5, SDXL, Flux training, LoRA merging, and DreamBooth mode
test/passes/diffusers/conftest.py	Test fixtures for mock models, accelerator, and test image folders
test/model/test_diffusers_model.py	Updated tests to mock is_valid_model for network-free validation
test/passes/diffusers/init.py	Empty init file for test package structure