* Fix performance regressions on OS 26 for pre-A19 Pro era devices; * Use Neural Accelerators for BF16 on A19 Pro / M5 devices; * Fix cold start slowness for Neural Accelerators; * LoRA import now respects your version selection more.

* Metal FlashAttention v2.5 w/ Neural Accelerators (preview) with breakthrough performance on M5 iPad / A19 Pro iPhones; * Support import AuraFlow models; * More compatibility improvement on importing Qwen Image LoRAs; * Some minor UI adjustments.

* Fix SDXL training on M3 / M4; * Support import Wan 2.2 5B models and LoRAs.

* Fix upload / download progress not showing properly. * Fix a crash upon first view wallet.

* Fix crash when using Qwen Image Edit 2509 without init image. * Support import Qwen image model. * Introducing "Boost" that can temporarily increase the Compute Units limit for a media generation task. * Add Bridge mode for gRPC server feature.

* Fix connectivity issues to Cloud Compute on iOS 26. * Add support for Qwen Image Edit 2509. * Add quantized BF16 support.

* Fix various UI glitches on iOS 26. * Fix image generation on iOS 26 for iPhone 12 / iPhone 12 Pro.

* All around of small improvements on onboarding flow; * Allow to resize the prompt box; * Show generation time for generated content.

* Fix crash on starting the app; * Fix Qwen Image Edit might not generate with good consistency issue.

* Support Qwen Image Edit for prompt-based image editing; * Support change brush size for eraser / paintbrush tool; * Further optimize Qwen Image to be faster for repeated runs.

* Add Privacy Pass support for Cloud Compute; * Improve decoding speed for Qwen Image generation; * Show sampler incompatible with certain models.

* Add Qwen Image support. * Fix MFA on 26 Beta. * Small UI polishes.

* Add support to specify LoRA for different stage of inference (Base / Refiner / All). * Fix MFA related crashes on 26 Beta.

* Added Wan v2.2 models to Official list; * Fixed various Refiner usability issues to facilitate Wan v2.2's High Noise Expert / Low Noise Expert separation; * Improved RAM usage especially for Wan model and when Refiner is on; * Various fixes for the keyboard issues on iPhone.

* Add support for CFG-Zero*; * Fix UniPC Trailing to make it work with flow-matching models. Also added UniPC AYS to make it competitive against DPM++ 2M AYS on SDXL / SD 1.5 cases; * Fix some cases after loaded image from files then using Kontext, it will wrongly recognized as inpainting mode; * Fix many small issues with the updated UI.

* Updated UI to streamline use of the app. * Add HiDream E1-1 to Official Models list.

* When using Cloud Compute, background the app during generation won't pause / fail the generation any more. * For iPad, there is a slightly updated visual for right side bar. * For "Causal Inference", added pad option allowing a frame to look at future frames for "causal inference". * Minor fixes on updating logic for order and filtering of API Providers.

* Support FLUX.1 Kontext [dev] model for image editing. * Fix import compatibility issues with some Wan 2.1 models. * Fine-tune with FLUX.1 Fill as base.

* Don't update settings after play video inline. * Fix crash when merging LoRA with FLUX base model in Model Mixing feature. * Slightly updated Compute Units UI. * Made the app available for devices below A12 (have to use Cloud Compute or Server Offload though).

* Support VACE with Wan 2.1 video models. Our support of VACE in this version enables you to use reference subjects for video generation and limited outpainting capability. * Support importing models that encode weights with FP8 E5M2 format. * Add "Uncurated" section in model selector.

* This is a hotfix to fix LoRA training issues with quantized weights.

* Add TeaCache support for HiDream series models. * Improved experience when using Community Configurations. * Improved experience when upload LoRAs for Cloud Compute. * ~2% performance improvements with DiT models when we selectively enable FP16 accumulators for GEMM. * Better chunking support when sending response back from gRPC server. * Always respect the seed set from script.

* Fixed support for Wan 2.1 CausVid LoRAs; * Added causal inference support for Wan 2.1 models; * Added support for "Bring Your Own LoRA" when using Draw Things+ Cloud Compute; * Fixed Chroma support with LoRAs; * Small polish on video related UIs; * Fixed Hunyuan LoRA import bug introduced when we added HiDream LoRA support.

* Added support to import HiDream LoRAs; * Added support to import Chroma (a variant of FLUX.1) models; * UI refresh for video generation support, now you can playback the video and save video more naturally; * Added experimental API model support.

* Better optimization for FLUX.1, Wan 2.1, Hunyuan, HiDream when running on 8GiB / 16GiB iP * Support import HiDream models; * Support create 8-bit models for Wan 2.1, Hunyuan, HiDream inside the app; * Show offline only icon for models that is not available on Cloud Compute; * When exporting video, it uses ProRes 444 format; * Some more minor UI clarifications.