Skip to content

v0.2 — image editing on GPU, size tiers, English UI

Choose a tag to compare

View all tags
@scsonic scsonic released this 24 Sep 00:14
· 16 commits to main since this release
v0.2
c3cceea

Superseded by v0.2.2 — its signature also differs from v0.2.2, so upgrading over it needs an uninstall first.

Needs the current models from evankuo/Qwen-Image-2.1-MNN — dit.mnn and dit.mnn.weight were re-exported on 2026-09-24. v0.1 does not work with the new files, and this release does not work with the old ones. Re-download both, or tap Download models in the app: it checks every file already on the phone against the repo's checksums (the new dit.mnn.weight has exactly the old size, so size alone would not catch it) and fetches only the ones that changed.

Changes

  • Image editing works on the GPU at every size. The DiT K/V cache is now one tensor per layer. As a single tensor, a Standard-size edit needed ~1 GB in one buffer, which is more than Adreno's 1 GiB limit, and every step read garbage.
  • Fixed the edit colours. The input image reached the VAE and the vision tower with red and blue swapped, which turned skin teal in the output.
  • Sizes: 7 aspect ratios × 3 pixel budgets (Standard ~512², Fast ~384², Tiny ~320²). The app shows the exact output size, and edit mode takes a size too.
  • Out-of-memory handling: each stage checks free memory first and shows a retryable Out of memory dialog. If Android still kills the app, the next launch says which stage and settings it was running. The VAE estimate is now fitted to phone measurements.
  • Prompt history, English UI, new README samples (all generated on a Snapdragon 8 Gen 2).

Measured on a Snapdragon 8 Gen 2, 20 steps

total
Text to image 448×576 (Standard) 451 s
Text to image 512×288 (Fast) 289 s
Text to image 320×320 (Tiny) 217 s
Image edit → 352×448 (Fast) 348 s
Image edit → 448×576 (Standard) 556 s

Assets: demo APK (debug-signed, arm64) and the qwenimage21 AAR. Built from c3cceea (MNN 1965a263).

Assets 4
Loading