# LumenX System Architecture Team **Black Rock** · Problem **P1: Robust Low-Light Image Restoration and Enhancement** · engine `core-v0.4.3` LumenX has two parts joined by one frozen interface ([`CONTRACT.md`](../CONTRACT.md)): - **`core/` (`lumenx-core`):** the processing engine: diagnosis, classical CV, deep models, hybrid fusion, metrics and report text. - **`server/` (`lumenx-server`):** the FastAPI service, upload handling, streaming, scheduling, evaluation harness and web UI. ## System diagram ```mermaid flowchart TD subgraph Client["Browser UI (server/lumenx_server/frontend)"] UP[Upload / gallery] --> VIEW[Live stages, before/after, maps, relight, candidates, metrics] BEN[Results & Performance tab] CAM[Live webcam] end subgraph Server["FastAPI service (server/lumenx_server)"] DEC[decode.py
type sniffing, EXIF, alpha, size limits, 12 MP downscale] LANE[lane.py
single GPU lane, bounded queue, cancellation] PUB[publish.py / runs.py
encode + atomic asset publication] SSE[POST /enhance · POST /enhance/stream SSE · GET /runs] META[GET /health /gallery /benchmarks /config] WS[WS /live] end subgraph Core["Engine (core/lumenx_core) · engine.py run() generator"] D[diagnose.py
illumination map, native-res noise σ,
SNR, hotspots, colour cast, severity] R[recoverability.py
recovered / estimated / unrecoverable] PX[CIDNet on 256-px proxy] T[tuner.py
12 candidates · CLIP-IQA + clipping/ΔE/edges
feasibility first, least-correction tie-break] C[classical.py
Retinex, highlight-protected gamma,
Lab CLAHE, partial gray-world] M[models/cidnet.py
HVI-CIDNet FP16, 512 tiles / 32 overlap
LOLv2-real or SID checkpoint by severity] F[fusion.py · repair_deep
clip + hotspot-ring mask, guided filter,
soft-knee highlight repair] N[denoise.py
gain-aware guided-filter denoise
inside repair / low-confidence regions] G[guardrails.py
audit: overexposure, chroma shift,
detail loss, halo] MR[metrics.py · report.py · render.py
pyiqa metrics, plain-English report, maps] Z[models/zerodce.py
Zero-DCE++ live + fallback] end UP --> SSE --> DEC --> LANE --> D D --> R --> PX --> T --> M --> F --> N --> G --> MR D -. priors: illumination, hotspots .-> F R -. confidence .-> N D --> C C -. classical mode / priors .-> F MR --> PUB --> SSE --> VIEW CAM --> WS --> Z M -. weights missing .-> Z EVAL[server/eval: evaluate.py → aggregate.py → mine_failures.py] --> RES[(server/results/benchmark_summary.*)] RES --> META --> BEN ``` ## Processing modes | Mode | Path | Used for | |---|---|---| | `classical` | diagnose → recoverability → classical → denoise → audit | classical-only baseline | | `dl` | diagnose → recoverability → CIDNet → audit (output untouched) | deep-learning-only baseline | | `zerodce` | diagnose → recoverability → Zero-DCE++ → audit | evaluation baseline, live webcam | | `hybrid` (default) | full path in the diagram | the proposed system | **Hybrid principle (deep-anchored):** the output equals the CIDNet output everywhere except inside a repair mask. The mask is built from classical priors (pixels CIDNet clipped at ≥ 0.98, plus overshoot rings around hotspots from the illumination map) and smoothed with a guided filter on input luminance. Inside the mask, the image blends toward a hue-preserving soft-knee highlight compression, with gain-aware denoising. Guardrails audit and report, but don't re-process (an automatic re-repair pass over-smoothed and was removed, see the ablation). ## Module map | Concern | File | |---|---| | Public interface, request/result types | `core/lumenx_core/contracts.py`, `contract_check.py` | | Orchestration, stage events, modes, ablation flags | `core/lumenx_core/engine.py` | | Diagnosis / recoverability | `diagnose.py`, `recoverability.py` | | Classical CV (Retinex, CLAHE, colour) | `classical.py` | | Gain-aware edge-preserving denoise | `denoise.py` | | Deep models, placement, FP16, tiling | `models/cidnet.py`, `models/zerodce.py`, `models/registry.py`, `models/runner.py` | | Hybrid fusion / highlight repair | `fusion.py` | | Adaptive candidate search | `tuner.py` | | Guardrail audit | `guardrails.py` | | Metrics (PSNR, SSIM, LPIPS, CLIP-IQA, MUSIQ, custom) | `metrics.py` | | Report text, map rendering | `report.py`, `render.py` | | All thresholds and weights | `core/lumenx_core/config/default.yaml` | | Upload decoding, limits | `server/lumenx_server/decode.py` | | GPU lane, queue, cancellation | `server/lumenx_server/lane.py` | | HTTP / SSE / WebSocket routes | `server/lumenx_server/app.py` | | Evaluation, aggregation, failure mining | `server/eval/evaluate.py`, `aggregate.py`, `mine_failures.py`, `run_final_eval.ps1` | ## Runtime - PyTorch 2.6, FP16 on CUDA (validated on an RTX 2050, 4 GB: peak 407 MiB at 1080p and 12 MP), FP32 on CPU. Device is chosen automatically. - Model fallback chain: CIDNet → Zero-DCE++ → classical. Every fallback is reported in the response `provenance`. - One engine lane: still-image requests are serialised; live webcam frames use a latest-frame slot and pause while a still request runs. - Deterministic for a fixed input, configuration and checkpoint on the same machine.