[Community] Add MagicTailor Personalization Training Script

### Model/Pipeline/Scheduler description

Recent advancements in fine-tuning techniques for text-to-image (T2I) personalization still struggle to distill visual concepts from reference images when there are both image-wide and spatially localized concepts present in each reference image.

The techniques in this paper are designed to improve component-controllable personalization, a novel task that pushes the boundaries of T2I models by allowing users to reconfigure specific components when personalizing visual concepts. This task is particularly challenging due to two primary obstacles: semantic pollution, where unwanted visual elements corrupt spatially localized concepts, and semantic imbalance, which causes disproportionate learning between the custom image-wide concept and spatially localized component concepts.

To overcome these challenges, MagicTailor leverages Dynamic Masked Degradation (DM-Deg) to dynamically perturb undesired visual semantics and Dual-Stream Balancing (DS-Bal) to establish a balanced learning paradigm for desired visual semantics.

### Open source status

- [X] The model implementation is available.
- [ ] The model weights are available (Only relevant if addition is not a scheduler).

### Provide useful links for the implementation

Paper: https://arxiv.org/pdf/2410.13370
Project Website: https://correr-zhou.github.io/MagicTailor/
Code: https://github.com/correr-zhou/MagicTailor
Contact: @Correr-Zhou

Start with the MagicTailor paper and project website, then inspect the linked implementation to understand the required training pipeline. Compare its components with the repository's existing training entry points; done means a usable MagicTailor personalization training script is integrated, with the unavailable model weights treated as an open limitation.

Vom Indexierungsmodell aus dem Issue-Text verfasst.