1. Architectural Design Goals
UpscaleX was engineered to eliminate the classic trade-off between zero-cost client privacy and heavyweight datacenter neural quality:
- Tier 1 — Client-Native Engine (
upscalex.video): Zero backend video transit, bounded O(1) streaming memory footprint via WebCodecs backpressure, and custom WGSL compute/fragment residual shaders executing at 15–60 fps on consumer GPUs. - Tier 2 — Serverless Ada Lovelace Cluster (
fast.upscalex.video): Cloudflare Pages Edge orchestration dispatching3×3edge-clamped multi-frame grids across auto-scaling 24 GB NVIDIA RTX 4090 / A5000 containers running FP16 23-RRDB Real-ESRGAN (RealESRGAN_x2plus,RealESRGAN_x4plus,RealESRGAN_x4plus_anime_6B) with zero idle GPU burn (workersMin: 0).
2. Solving Cloud Per-Frame Dispatch Overhead via 3×3 Edge-Clamped Montages
Naively sending 300 individual video frames over 300 separate REST requests incurs TLS/HTTP framing latency and container queue contention. Conversely, uploading raw multi-hundred-megabyte video blobs to centralized storage buckets introduces egress delays and privacy liability.
UpscaleX's cloud engine solves both problems with 3×3 Padded Frame Montages:
- Up to 9 sequential frames are packed into a single composite grid canvas (capped at 640 px long-edge per cell → 1944 px composite width).
- Each cell is surrounded by a 4-pixel reflected/clamped border gutter so 3×3 convolutional receptive fields never bleed dark border pixels across adjacent frames.
- Up to 3 montage jobs execute concurrently in ~1.1–1.5 seconds per grid on RTX 4090 Tensor Cores—achieving an effective throughput of up to 18–25 upscaled frames per second after container warm-up.
- Returned 2×/4× grids are cropped precisely inside their scaled gutter coordinates and fed straight into hardware
VideoEncoder(`avc1.640033`, High Profile Level 5.1) alongsideAudioEncoder(`mp4a.40.2` AAC) into an ISO-BMFF fast-start MP4 container stored in IndexedDB.
3. Cryptographic Billing Integrity Without Centralized User Media Databases
Fast Cloud credit balances are reconciled directly against authoritative payment provider ledgers via Cloudflare Edge Functions and signed with HMAC-SHA256, ensuring tamper-proof pay-as-you-go metering ($0.05–$0.12/min) with automatic client-side rollback if any GPU batch is cancelled.