* Add support for import and run SD3 / PixArt models with LoRA;
* Fix quality issues with PixArt models when running on A12-A16, M1-M2 GPUs.

* Support PixArt Sigma XL models. These are lightweight models punch above its weight.
* Support importing SD3 medium, PixArt Sigma XL fine-tunes in the app.
* Fix issue with importing SDXL LoRAs that fine-tuned over diffusers format OpenCLIP G won't be recognized.
* Allow specify separate CLIP text for SD3 medium models.
* Smaller RAM usage for SD3 medium when MFA is disabled.

* Fix issue that SDXL cannot be imported any more in previous version.
* Improve compatibility with fine-tuned PixArt Sigma models.

* Fix issue when remove a model downloaded from Community Models, it won't show up back in the Community Models list.

* Add Stable Diffusion 3 Medium support. The model should be available under "Community Models" section. It should be compatible with devices has 6GiB or more RAM. It is best to be used with "* Trailing" samplers.
* Changed how CLIP Skip works for SDXL models. Now CLIP Skip 2 should work similarly to CLIP Skip 1, hence for Pony models, even if you use CLIP Skip 2 would generate reasonable images.
* Added "DDIM Trailing" and "DPM++ 2M Trailing" samplers.
* Support more variant of Stable Cascade fine-tuning importing.

* Fix crash related to importing Scribble / SoftEdge ControlNet not having preprocessor set properly.
* Fix crash related to negative embedding longer than 77 and have mis-match handling between two tokenizers.
* Fix issue that the "Use CoreML for LoRA" option is not surfaced properly in Machine Settings.
* Fix crash when using acceleration LoRA with CoreML.
* Fix crash when supplying init_images for HTTP API will crash when the image doesn't match the size on parameters.

* Support importing fine-tuned Stable Cascade models.
* Fix SDXL export issues to be more compatible with other SDXL clients.
* Fix slider in script-driven UI display issues where certain floating values cannot be used.
* Add new APIs for script that can manipulate moodboard.
* Fix issue that full ControlNet SDXL not imported correctly.

Hotfix issues regarding script related widgets in previous release.

1. Add InstaStyle support. When select IPAdapter for either SD v1.5 or SDXL models, you can now choose whether to borrow only style, style & layout, or borrow everything from the reference images.
2. Added "Align Your Step" based schedule for samplers. You can generate good quality images from vanilla non-accelerated models with as little as 10 steps when using "DPM++ 2M AYS" sampler.
3. Fixed bugs: when copying image in with settings, it should recognize samplers correctly; when doing LoRA merging in model mix flow, it now does that at higher precision (FP32); Now when using accelerated LoRAs / models, you can select all samplers without all fallback to LCM sampler.

* When doing model mix, now can add LoRAs into the mixed model.
* Fix more software keyboard usability related issues.
* Fixed LoRA training UI image size picker issue.