Skip to content

Tiggy-Chan/LocalMotion

Repository files navigation

LocalMotion

LocalMotion is a clean-room Android prototype for on-device image-to-video generation on Snapdragon 8 Gen 3 devices.

Current State

  • New Android project rooted at /app
  • Native sidecar backend rooted at /native
  • Runtime bundle installer with HTTPS download, SHA-256 verification, and atomic swap into files/models/video-v1/
  • Native sidecar /health now probes runtime bundle completeness and reports whether the APK can enter QNN mode
  • First-pass SD1.5 conversion scaffolding under tools/convert-qnn/sd15/
  • SD1.5 checkpoint has already been converted to:
    • models/shortlist/converted/sd15-diffusers
    • models/shortlist/converted/sd15-onnx
  • Placeholder local fallback pipeline for:
    • image import and square crop
    • prompt-driven lightweight stylization
    • staged SSE progress reporting
    • local AVC/MP4 encoding

Repository Layout

  • app/: Compose UI, services, local client, artifact storage, export/playback
  • native/: sidecar process with /health, /generate_clip, /cancel
  • tools/convert-qnn/: SD1.5 export helpers and runtime manifest packager
  • docs/architecture.md: implementation notes and next integration steps
  • docs/sd15-first-npu.md: first real NPU milestone for txt2img / img2img
  • docs/wsl-sd15-qnn-checklist.md: WSL-side setup checklist before QNN work

Important Notes

  • This repository does not ship QNN SDK assets or model weights.
  • The app can now install and validate a real runtime bundle, but the native inference core is still gated by Qualcomm SDK-bound code that is not committed in this repository.
  • The first realistic NPU milestone is SD1.5 txt2img/img2img, not video generation.
  • Set localmotion.runtimeManifestUrl=<https-url> in local.properties to bake a default runtime bundle manifest URL into the APK.

About

On-device AI Video Generation for Android. Text-to-Image, Image-to-Video, Offline.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors