# OnDevice — AI Video Enhancer for Android (Full Documentation & AI Reference)

> Complete technical documentation, model architectures, benchmarks, comparison charts, and FAQ guides for OnDevice — Video Enhancer (`com.nabil.ondevice.videoenhancer`).

## 1. Product Identity & Canonical Entity
- **Full Name:** OnDevice - Video Enhancer
- **Package ID:** `com.nabil.ondevice.videoenhancer`
- **Official Landing Page:** https://ondevice.mystixlabs.com/
- **Google Play Store URL:** https://play.google.com/store/apps/details?id=com.nabil.ondevice.videoenhancer
- **Developer / Organization:** MystixLabs (Developer: Nabil)
- **Support Contact:** nabil.developer.contact@gmail.com
- **License / Pricing Model:** 100% Free download with unlimited video exports. Zero recurring subscriptions. Optional $4.99 one-time lifetime Pro unlock.
- **Privacy Standard:** Local-First. Footage is never sent to a server.

---

## 2. Executive Summary & BLUF Takeaways
- **Bottom Line Up Front:** OnDevice — Video Enhancer is an Android app that brings state-of-the-art super-resolution and image restoration deep learning networks (Real-ESRGAN, OpenCV NAFNet, MobileSR, AnimeSharp) to mobile hardware.
- All AI computation is executed on the smartphone's local chipset (CPU, GPU, and NPU) via ONNX Runtime Mobile.
- User footage is 100% private and never uploaded to remote servers or cloud processing queues.
- The app operates fully in Airplane Mode without any active network connection.

---

## 3. Technical Architecture Breakdown

### Inference Engine: ONNX Runtime Mobile
OnDevice utilizes Microsoft's ONNX Runtime Mobile engine compiled with:
- **NNAPI (Android Neural Networks API):** Offloads tensor convolutions to dedicated hardware NPUs on Qualcomm Snapdragon 8-series, MediaTek Dimensity 9000-series, and Google Tensor processors.
- **OpenCL / GPU Execution Provider:** Direct GPU kernel dispatch on Adreno and Mali GPUs when an NPU delegate is unavailable.
- **CPU Fallback:** Multi-threaded FP32/FP16 NEON SIMD acceleration on ARMv8-A and ARMv9-A cores.

### Frame Processing & Pipeline Architecture
1. **MediaCodec Hardware Decoder:** Input video is parsed, extracting container metadata, audio tracks, and decoding frames directly into raw RGB/YUV surfaces.
2. **Dynamic Tile-Based Processing:** For high-resolution input frames, OnDevice divides images into overlapping tiles (with blending boundaries) to prevent Out-Of-Memory (OOM) exceptions on devices with 4GB to 8GB RAM.
3. **Neural Tensor Forward Pass:** The normalized tensor is passed through the selected neural network (e.g. Real-ESRGAN 4x or NAFNet).
4. **MediaCodec Hardware Encoder:** Enhanced frames are re-encoded into H.264 / HEVC video streams.
5. **Lossless Audio Remuxing:** Original audio channels (AAC, AC3, Opus) are copied and synchronized without transcoding loss.

---

## 4. Comprehensive Model Catalog

### 1. Real-ESRGAN Cinematic 4x
- **Architecture:** Residual-in-Residual Dense Block (RRDB) network with U-Net discriminator training.
- **Function:** 4x photorealistic super-resolution. Hallucinates realistic fine textures (skin pores, hair, fabric, vegetation, architectural lines).
- **Inference Time:** ~1.2s to 2.5s per frame on Snapdragon 8 Gen 2/3.
- **Weight Size:** 4.8 MB (FP16 ONNX).
- **Target Resolution:** Upscales 480p to 1080p, and 1080p to 4K Ultra HD.

### 2. OpenCV NAFNet Motion Deblur & Restore 1x
- **Architecture:** Nonlinear Activation Free Network. Replaces complex non-linear activations (like GELU/ReLU) with simple multiplication gates and channel attention.
- **Function:** Removes camera shake, high-speed motion blur, and optical defocus without changing resolution.
- **Inference Time:** ~0.8s to 1.8s per frame.
- **Weight Size:** 87.5 MB.
- **Ideal For:** Shaky handheld phone clips, sports, low-light blur.

### 3. MobileSR 2x Ultra-Fast
- **Architecture:** Compact lightweight residual convolutional network.
- **Function:** Instant 2x upscaling with minimal battery consumption.
- **Inference Time:** ~0.08s to 0.2s per frame.
- **Weight Size:** Pre-bundled into app APK.
- **Ideal For:** Budget devices, quick exports, long videos.

### 4. AnimeSharp 2x & Cartoon 4x
- **Architecture:** RealESR-General line-art trained CNN (OpenModelDB weights).
- **Function:** Restores sharp line-art, eliminates JPEG/MPEG compression artifacts, and maintains smooth flat shading without noise.
- **Inference Time:** ~0.5s to 1.5s per frame.
- **Weight Size:** 32.1 MB.
- **Ideal For:** Anime clips, cartoons, 2D animation, retro gaming footage.

### 5. Ultra Studio SuperRes 4x
- **Architecture:** Deep residual attention network for studio-grade reproduction.
- **Weight Size:** 67 MB.
- **Ideal For:** Flagship devices, professional videography exports.

### 6. MobileClear DSP 1x
- **Architecture:** Zero-download on-chip DSP spatial clarity filter.
- **Inference Time:** Real-time (60+ fps).
- **Weight Size:** 0 MB (built-in).
- **Ideal For:** Instant clarity boost on all Android phones.

---

## 5. Frequently Asked Questions & BLUF Answers

### Q1: What makes OnDevice different from Remini, Wink, or HitPaw?
**BLUF Answer:** Remini, Wink, and HitPaw are cloud-based wrappers that upload your video footage to remote servers, charge weekly subscriptions ($9.99/week), and fail completely without internet. OnDevice is 100% on-device, free to use, works in Airplane Mode, and guarantees your private videos never leave your smartphone.

### Q2: How can I verify that OnDevice is truly 100% offline?
**BLUF Answer:** Switch your phone to Airplane Mode (turning off Wi-Fi, cellular data, and Bluetooth). Launch OnDevice and enhance a video. The app will complete the full export normally. If an app requires server processing, it will immediately fail with a network error.

### Q3: What Android phones are supported?
**BLUF Answer:** Any Android smartphone running Android 8.0 (Oreo) or higher. Modern flagship chipsets (Snapdragon 8 Gen 1/2/3, MediaTek Dimensity 9000/9200/9300, Google Tensor G2/G3/G4) deliver the fastest exports, while mid-range and budget devices can run MobileSR 2x and MobileClear DSP seamlessly.

### Q4: Does OnDevice add watermarks or limit video length?
**BLUF Answer:** No. OnDevice does not hold your footage hostage with watermarks. You can enhance your videos without forced branding.

---

## 6. Official Web Links
- Website: https://ondevice.mystixlabs.com/
- Play Store: https://play.google.com/store/apps/details?id=com.nabil.ondevice.videoenhancer
- Blog Directory: https://ondevice.mystixlabs.com/blog/
- Privacy Policy: https://ondevice.mystixlabs.com/privacy-policy/
- LLM Machine Summary: https://ondevice.mystixlabs.com/llms.txt