v2.8.1

Device link

You can now run models on another device over your local network. On the device that has the models, open Device link from the menu and enter host mode; on a second device, connect to the address it shows. The second device lists the host's models and runs generation, tokenize, and upscaling through it — no need to download the models again, and the controlling device doesn't even need a Qualcomm chip.

• Host mode keeps the screen awake while it serves, with an optional pure-black screen to protect OLED displays.

UltraFix

• UltraFix can now start from a local image. Long-press the upscale or UltraFix button on the result page to import one; resize it to fit, and smaller images can be upscaled first, then repaired.

Variants :

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

Co-Authored-By: Claude Opus 4.8 & Claude Fable 5

v2.8.0

Anima models

Added support for Anima models (developed by Circlestone Labs). Like SDXL, they run on Snapdragon 8 Gen 3 and newer devices.

• Anima models currently support the same features as SDXL models, except UltraFix, which remains available only for SDXL DMD2 models.
• Because Anima models are slow, only the Turbo version is supported for now (fused with the official Turbo LoRA). Anima models are not shown in the built-in download list for now — you can download and import it from https://huggingface.co/xororz/anima-qnn.

Important

If your device has 12GB of RAM and Anima models doesn't run with the default settings, enable DiT sequential loading in Settings. This makes each denoising step much slower, so if Anima already runs without it, leave it off.

Upscaler

• Added 2x / 3x / 4x scale options. Note that the model always upscales by 4x; 2x and 3x simply downscale the 4x result for other purposes.

Variants :

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

Co-Authored-By: Claude Opus 4.8

v2.7.0

UltraFix

On-device 4K high-quality anime image generation

• UltraFix — a new tiled-diffusion repair pass for upscaled images. After upscaling a result, tap UltraFix to re-run the model over the image tile by tile and restore real fine detail at high resolution, instead of just a smoother enlargement. Available only on SDXL DMD2 models (it runs at CFG 1 / few steps). It has its own Steps / Denoise controls (defaults 10 / 4): keep denoise-steps ÷ steps ≤ 0.5. Use it only on anime models — on realistic models it tends to duplicate subjects. Use it as a finishing pass on images you like, and run it again on the result to build up more detail. (Based on the PixelRush algorithm.)

Generation

• The back gesture during generation now interrupts the run instead of exiting the backend — it stops within about one step and frees the NPU, so you can start again right away.
• Improved inpaint blending when compositing the result back into the original image, removing the faint square seam.

History & Results

• Added Favorite for generation results, with a favorited / not-favorited filter in history.
• Added Upscale and UltraFix buttons to history images.
• The result page now shows the most recent image when the current model has history, instead of an empty state.
• Added a global history screen covering every model — including models you've deleted.
• Deleting a model now keeps its history by default, with an optional checkbox to also delete the history data.
• History now loads as a paged query, so large histories stay smooth.

Models

• Support pinning models to the top and renaming custom models (history, parameters and pin state all move with the new name).

App & Settings

• Added history import / export — back up your whole history (images + parameters) to a zip and restore it later or on another device.
• Added a manual Clean Temp Files action to clear temporary files that in some cases can't be cleaned up automatically (e.g. an interrupted download or extraction).

Fixes

• Fixed blurry preview thumbnails on the result page.
• Fixed the scroll-position memory of the recent-results strip on the result page.
• Fixed non-square images not exiting preview mode when tapping outside the image.
• Fixed a crash on some devices when the upscaler page rendered an overly large image.
• Fixed the upscaler download not following a changed download source, and the upscaler existence check; downloads are now integrity-checked.
• Fixed a "job cancelled" error that sometimes appeared even though the model downloaded successfully.
• Fixed backend lifecycle issues during model exit / load that had existed since the MD3E UI update.
• Other i18n and performance refinements.

Variants :

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

Co-Authored-By: Claude Fable 5 & Claude Opus 4.8

v2.6.4

• Replaced the built-in SDXL Base model with CyberRealistic v10, and added built-in DMD2 variants of Illustrious v16 and CyberRealistic v10 that generate in just a few steps — several times faster than the regular versions.

• Converted models can now specify their own default parameters by adding a config.json inside the zip (all keys optional):

  {
      "default_prompt": prompt,
      "default_negative_prompt": negative_prompt,
      "default_scheduler": scheduler_id,
      "default_steps": steps,
      "default_cfg": cfg
  }

  After importing, the model's default parameters come from config.json. All DMD2 models on https://huggingface.co/xororz/sdxl-qnn already have their defaults set to:

  {
      "default_scheduler": "lcm",
      "default_steps": 8,
      "default_cfg": 1
  }

• Added in-app links to pre-converted models and the documentation.

• Fixed upscaled inpaint results not being stitched back into the original image when saving.

Variants :

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

Co-Authored-By: Claude Fable 5

v2.6.3

• The tag autocomplete popup now has an action toolbar: add a new tag, clear the current one, raise or lower its weight in 0.1 steps (it understands the (tag:1.1), (tag) and [tag] forms), undo/redo your prompt edits, and close the popup. The popup now anchors just above the line you're typing and follows the keyboard so it never covers the cursor.
• Prompt text beyond the 77-token CLIP limit is greyed out — down to the exact character where it gets cut off — so you can see what the model actually reads.
• Fixed result-page thumbnails occasionally appearing blurry.
• Fixed flickering when switching the displayed image.
• Fixed square corners appearing on cards and images during their animations.
• Fixed several prompt autocomplete popup behavior issues.

Variants :

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

Co-Authored-By: Claude Opus 4.8

v2.6.2

• Faster model loading via mmap across the board — the CPU/NPU runtime models, the token-embedding tables, and SD 1.5's per-resolution patch step — for a small load-time improvement in all cases and a noticeable reduction in generation time in SDXL lowram mode.

Variants :

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

Co-Authored-By: Claude Opus 4.8

v2.6.1

• Tag autocomplete now does fuzzy matching for both the tag dictionary and the translation dictionary: type an abbreviation or skip characters (e.g. blnht → blue_night, or search by translation) and it still finds the tag, with the matched characters highlighted. It also stays fast on large dictionaries.
• Tags containing parentheses are now auto-escaped on insertion (e.g. aqua_(konosuba) → aqua \(konosuba\)), so they aren't parsed as prompt weighting.
• Added Korean localization (#219).

Variants :

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

Co-Authored-By: Claude Opus 4.8

v2.6.0

Local Dream ❤️️ Material 3 Expressive

• App-wide pass moving to Material 3 Expressive. Added a theme color setting. Improved predictive back gesture support.
• Added "Allow LAN access" in settings — when enabled, the backend listens on 0.0.0.0 instead of 127.0.0.1, and CORS is now allowed so you can write browser scripts to access it.
• Fixed a gap in the app icon.

Variants :

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

Co-Authored-By: Claude Opus 4.7

v2.5.3

• Fixed SDXL's timestep spacing configuration: it should be "trailing" rather than "leading". This slightly improves image quality — note that the same seed may produce slightly different results from before.
• Improved the CLIP cache. Prompts are now hashed and persisted to disk, meaning a prompt you've used before will never be re-encoded a second time — about 1 s faster per SDXL run. You can clear the cache from the File Manager in settings.
• Reproduce dialog now lets you tick which parameters to copy back instead of an all-or-nothing prompt; both share and reproduce dialogs gained a select-all / deselect-all toggle.

Variants :

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

Co-Authored-By: Claude Opus 4.7

v2.5.2

• Aspect ratio for SDXL. Generate non-1:1 images on SDXL models via preset chips (1:1 / 3:4 / 4:3) or custom W:H, supported across txt2img, img2img and inpaint. The long axis is fixed at 1024 and the short axis is scaled from the ratio; under the hood it runs as an inpaint over the 1024x1024 canvas, so generation time matches a normal 1024x1024 run.
• Show model id in result and history parameter details.
• Include model id in shared parameters, and add a clipboard import button in advanced settings.
• Clearing the clipboard now also happens when the import dialog is cancelled.
• Fixed the generation timer reporting a shorter time than the actual run.

Variants :

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

Co-Authored-By: Claude Opus 4.7