v2.3.0

This release features a major refactor of the NPU HighresPatch functionality, reducing disk usage while adding support for more resolutions.

• Added support for 512x768, 768x512, 768x1024, and 1024x768 resolutions (Only for 8gen1 and above).
• Patched high-resolution models are now retained directly in memory, reducing disk space consumption.
• Moved the HighresPatch button from the Model Card to the Model Run interface, so switching resolutions no longer requires exiting the page.
• Enabled img2img and inpainting modes for all resolutions. Note that changing resolutions which alter the aspect ratio will clear the current input image. For resolutions >768x768, it is recommended to generate a low-res image first, then use img2img with a denoise strength of 0.75 for Highres fixing.
• Removed support for SD2.1 due to poor performance and the overhead of adapting new features. Future maintenance will focus exclusively on the SD1.5 model series.

Other Improvements & Bug Fixes

• Optimized image loading in the history view and removed the 100-item display limit.
• Stored token_emb in fp16, reducing memory and disk usage by approximately 70MB.
• Mapped built-in model downloads for 8sGen4 and 8Gen5 devices to _8gen2, allowing these devices to utilize _8gen2 models for significant speed improvements.
• Refactored the download function to use a single ZIP archive instead of multiple files. Implemented a foreground service to allow switching to other apps during the download process.
• Added padding to the image crop interface to prevent conflicts between image edges and system navigation gestures.
• Fixed calculation errors when using embeddings or weighted prompts.
• Fixed a vertical scrolling issue on small screens when content overflows.
• Fixed an issue where models were forced to reload when page dimensions changed.
• Custom models are now sorted alphabetically in the display list.
• Optimized various UI style details.

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.

v2.2.1

• Fixed an issue where manually switching UI during batch generation would interrupt the generation process
• Fixed an issue where buttons were covered by images on screens with large aspect ratios
• Model names are now automatically extracted from filenames during model import (#116)
• Improved UI switching behavior

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.

v2.2.0

• Support single-click serial batch generation feature (#111)
• Added generation history feature for each model, displaying only the most recent 100 images
• Display the most recent 20 historical generated images below the image generation results
• Support copying generation parameters from historical images to current parameters
• Fixed a memory leak issue caused by repeated upscaler runs

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.

v2.1.2

• Add draggable zoom feature for super-resolution page
• UI performance improvements.

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.

v2.1.1

• Add a standalone super-resolution page to upscale local images (NPU Only). A shortcut is available in the top-right corner for Qualcomm devices.
• UI performance improvements.

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.

v2.1.0

• Add built-in upscaler, supporting NPU models only. It uses realesrgan_x4plus_anime_6b for anime images and 4x_UltraSharpV2_Lite for realistic photos.
• Images with a width or height greater than 1024 will now be saved as JPEG (e.g., upscaled or inpainted images) for better performance. Smaller images will remain in PNG format.
• Improve app UI responsiveness after image generation is complete.
• Fix occasional UI freeze when modifying prompts or generation parameters.

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.

v2.0.0

v2.0.0 is now available.

• Support for using parentheses to change prompt weights, with the same format as WebUI's standard, for example (best quality:1.5).
• Support for importing embedding files in Settings, such as EasyNegative. Only safetensor format is supported. please convert pt format to safetensor format before importing, you can use this script to convert. Note that the trigger word for the embedding is the same as your imported filename, case-insensitive.
• To support the above two features, I have refactored the text encoder format. v2.0.0 is compatible with the old text encoder format, so old models can still be used, but old encoder files cannot use the above two features. Therefore, for app preset models, after updating you need to download the clip_v2.mnn file. For custom CPU models, you need to reconvert them in the app. For custom NPU models, I have updated the NPU conversion script which will convert the new text encoder. For precompiled models on HF, I will gradually update them to clip_v2 in the future.

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.

v1.9.1

• The output in inpainting mode is now stitched with the original image upon saving.
• Update blending algorithm to Laplacian Pyramid in inpainting mode.
• Optimize build config for android apk.

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.

v1.9.0

Happy Mid-Autumn Festival!

• Update QNN runtime libs from v2.29 to v2.39. Now Snapdragon 8 Elite Gen5 is supported for NPU generation.
• Fix issues with Dimensity chip GPU runtime failing to utilize GPU and producing incorrect results.
• Special thanks to Mr.J for pre-compiling a lot of NPU models. You can find NPU models at xororz/sd-qnn and Mr-J-369.
• Download _min if you are using non-flagship chips. Download _8gen1 if you are using 8gen1. Download _8gen2 if you are using 8gen2/3/4/5.

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.

v1.8.4

• Add support for importing custom npu models via a zip file. Follow the guide here to create custom npu models.
• You can get some models here. For RealisticVisionHyper, cfg1.5 steps 6 is recommended.
• Download _min if you are using non-flagship chips. Download _8gen1 if you are using 8gen1. Download _8gen2 if you are using 8gen2/3/4.

Variants :

• LocalDream_xxx_with_filter: Block NSFW Results. Same as Google Play version.
• LocalDream_xxx: No filter.