E-OS v4 — Full-Frame Generative OS (NPU)

Version 0.4 of E-OS: the ENTIRE 384x832 UI frame is generated by one model (FrameGenerator, 0.88M params) on the Hexagon NPU of a Samsung Galaxy S26 Ultra. No masks, no cell gating — the code-drawn rough layout (192x416) is only a conditioning hint ("was liegt wo an"), everything visible is generated.

APK (install over v0.3.x — same signature)

apk/e-os-npu-v0.4.apk · versionCode 5 · package com.killerboss.eos Direct download: https://huggingface.co/KillerBoss/e-os-v4/resolve/main/apk/e-os-npu-v0.4.apk

Highlights

  • 8 screens: Lock (Aurora), Home (Glühwürmchen), App-Drawer, Rechner, Wetter (Sonne/Wolken/Regen/Sterne), Musik (Spektrum-Visualizer), Einstellungen (Akzentfarben + KI-Toggle + Diagnose), Info
  • Zurück-Buttons: in-UI (oben links), Nav-Bar (Zurück/Home/Apps), Hardware-Back
  • Slide-Übergänge, Ripples auf jedem Tastendruck, 60-HZ-Ziel (NNAPI -> Hexagon, sauberer CPU-Fallback, Code-SAFE-MODE — startet immer)
  • Diagnose-Anzeige im OS (Settings > Diagnose, liest crash.txt)

Model

params 0.88M
MACs (app res) 3.9 G
input hint 1x3x192x416 + state 1x32
output frame 1x3x384x832 (sigmoid)
ops Conv/SiLU/FiLM/Concat/Resize/DepthToSpace — NNAPI-friendly

Dataset: 3,200 samples (1,600 scenes x 2 Zeit-Varianten), hint->styled pairs, state-Vektor 32-dim (Screen, Übergang, Ripples, Tageszeit, Hue, Musik, Wetter).

Train

python scripts/gen_os4_dataset.py && python scripts/train_os4.py (chunked, resume via checkpoints), then export_os4.py -> ONNX opset 17, ai_hub_os4.py -> TFLite + NPU profile on Galaxy S26 Ultra.

Repo layout

  • apk/ · app/ (Kotlin-Quellen) · scripts/ (komplette Pipeline)
  • models/ (ONNX) · npu/ (TFLite + Profile) · eval/ (PSNR + Previews)
  • dataset/ (states, scenes, hint-PNGs, Frame-Shards; frames.mem = 3200x832x384x3 uint8)
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support