Pocket production studio: HTML motion clips -> pixel-exact 16:9 and 9:16 MP4s, and scripts -> voice audio (Qwen3-TTS or Breeze2-TTS). Renders on this phone or on a Kaggle GPU farm. Zero third-party code; no screen capture, no ffmpeg.
Mission-control identity: steel panels on a blue-cast void, warm chalk ink, two semantic accents (AMBER = PIXEL, TEAL = VOICE), Chakra Petch display / Saira body / IBM Plex Mono data. The pipeline rail is the signature element — the render sequence, its state, and its captured decisions in one strip. Typefaces bundled as assets.
Push to GitHub. The apk workflow (.github/workflows/build.yml) builds app-debug.apk with Gradle 8.9 + AGP 8.5.2 and uploads it as the app-debug artifact. Download it from Actions -> latest run -> Artifacts, then install on the phone (allow "install unknown apps" for your browser/files app).
The repo commits debug.keystore, so every rebuild is signed identically and installs over the previous version without uninstalling.
The home screen is two engine cards — PIXEL (video) and VOICE (audio), each in its engine accent — plus a machine footer (queue/farm counts) and one door to the Machine room: farm setup, farm jobs, overnight queue, generation prompt, all out of the creative path.
PIXEL flow — the pipeline rail (SOURCE · FRAME · RENDER · EXPORT, carrying each step's captured decision) persists across every screen:
- SOURCE - Choose HTML (system file picker,
text/html) or paste code. - FRAME - preview with AUTO (amber brackets lock the DOM-detected 16:9 or 9:16 frame) or MANUAL framing (pinch/drag zoom+pan, editor-style CROP with corner handles + thirds grid, FULL/CENTER presets). RATIO 16:9/9:16 (swaps output width/height), Resolution 480p/720p/1080p/4K, FPS 24/30/60, BITRATE, ENHANCE, SFX (when the page declares #sfx events) and editable duration (5-600 s) live on this screen.
- RENDER - progress bar with
frame N/T - r f/s - ETA; CANCEL aborts cleanly (encoder released, partial file deleted). - EXPORT - file card read back with MediaMetadataRetriever: name, size, duration, resolution, VERIFIED/MISMATCH line. OPEN / SHARE / RENDER ANOTHER.
Output lands in Movies/NexonStudio/NexonStudio_<stamp>.mp4, with Pictures/NexonStudio/<stamp>_frame0.png for eyeball QC (MediaStore forbids images under Movies/).
Zoom and crop are applied as a CSS transform on the page itself, so content re-rasterizes at the output resolution - crisp text and vectors at any zoom, never a bitmap upscale of the capture. Known limits: position:fixed elements and vw/vh-sized layout will not pan; raster images soften when zoomed. Conforming HTML (a .fit 16:9 or 9:16 frame containing .stage) still auto-detects; non-conforming HTML falls back to manual framing instead of aborting, initialized by an automatic content-bounds fit (contain + center) so phone-authored pages fill the frame instead of rendering small in the void. The one hard requirement: at least one CSS animation - the whole engine scrubs document.getAnimations().
A color grade (contrast ~1.12 around mid-gray + saturation 1.18) applied as a root CSS filter inside the GPU draw - identical math to the previous native ColorMatrix, still at native speed. It is deliberately NOT an AI model - the app stays zero-dependency and nothing ever leaves the device. frame0.png is exported with the same grade so QC still matches.
- The video frame is
<div class="fit">(CSSaspect-ratio: 16/9, or9/16for vertical - generation prompt Part 4B; contains.stage);#video-frameis also supported. Everything outside it is crop chrome;?headless=1(plus an injected class) hides it. - All motion is CSS animations; the timeline is scrubbed via
document.getAnimations()+currentTime. - Webfonts: every family requested via the page's Google Fonts
<link>is checked (default pairing Anton + IBM Plex Mono; deviations allowed). A requested font that has not loaded stops preparation instead of silently rendering narrower fallback glyphs; retry once it is available, then WebView caches it. - Page void background:
#0A0C10(the renderer's canvas pre-fill matches it). - The exact injected JS lives in
app/src/main/java/com/brollrender/app/JsContracts.kt(frozen strings).
- Never captures the screen: offscreen WebView, hardware layer, drawn manually per frame - GPU-direct into the encoder surface when a per-render probe validates it, CPU readback fallback otherwise.
- Never upscales: AUTO framing must match the target within 2 px or fall back to manual; manual zoom/crop re-rasterizes the page instead of resampling pixels.
- Corner detection is DOM geometry (the page reports its own rect), never vision.
In-app voice generation on a Kaggle GPU notebook — same one-command pattern as the render farm:
- New Kaggle notebook, Accelerator = GPU (T4 x2).
- In the app: Voice studio → Start the voice server → Copy command, paste it into a Kaggle cell and run. It installs everything (flask, qwen-tts, cloudflared), starts the server plus a Cloudflare tunnel, and prints a BASE URL in a READY block.
- Paste the BASE URL into Voice studio → Connection, save, test.
- Three modes: Custom (9 preset speakers + instruct for tone/emotion), Clone (3–15 s reference audio + its exact transcript), Design (describe the voice in natural language). Paste the script —
{here complete voice over}marks narration start — and generate. - The finished WAV lands in
Music/NexonStudio/with play/share buttons.
Server files live in qwen3tts/ (run.sh one-command launcher + server.py); Qwen3-TTS models lazy-load per mode on first request. Re-running the cell is safe: it kills the old server/tunnel and starts fresh.
A second TTS engine: BreezeBlue/breeze-tts-2 via breezeblue-ai/breeze-tts — top-ranked open-weights TTS on the Artificial Analysis voice arena (EN + ZH, ~3B, eager attention, ~7.7 GiB VRAM: fits one T4, which runs its hard-coded bf16 via emulation — dtype fallback documented in run.sh). Same one-command Kaggle pattern, same job-queue API, its own port (5001 — never clashes with a qwen server on 5000):
!rm -rf /kaggle/working/V && git clone --depth 1 https://github.com/shivaww/Recorder.git /kaggle/working/V && cd /kaggle/working/V/breeze2tts && bash run.sh
Modes: custom (plain text; optional natural-language instruct — Breeze has no preset speakers, so speaker is ignored), clone (reference audio + its exact transcript, mandatory), design (describe the voice in natural language), direction (clone + delivery instruction). Inline emotional tags — (sigh), (laugh), [笑] — are interpreted in the text itself. Output: mono 24 kHz WAV. Note: model weights carry a research/non-commercial license (code is Apache-2.0).
The app's Connection field works with either engine's BASE URL (verified: custom/clone/design flow through unchanged; in custom mode the instruct line describes the voice) — same endpoints and form fields as the qwen server (/health, /generate, /status/<id>, /download/<id>), so the existing client contract works unchanged.