Application Size Optimization (Modular AI Model Management)

[rohan-pandeyy]

Eliminate bundled ONNX models (~348 MB) from the installer. Replace with a post-install download system that is hardware-aware, integrity-verified, and user-controllable.

An in-depth explanation of each of these implementations can be found here

---

Pre-implementation

• Upload all non-Ultralytics ONNX models to models-v1.0 GitHub Release tag
• Compute and record SHA-256 checksum for each uploaded file

Computed SHA-256:

FILE	SHA-256 SUM
FaceNet_128D.onnx	c37946ea8cce94141777dcbcccb8a61786c9a7c4f0c9f40471bd27029fa664ed
YOLOv11_Medium.onnx	f53a0bf6a141788f329ab92b438045f0ebac73ff10285b9ef0d06551cdbd01ea
YOLOv11_Medium_Face.onnx	fd3ab085d776ee1ce59fd47def82c2abf49bbfde3a10a1b5b660b76cb41a6912
YOLOv11_Nano.onnx	634279b40c07c6391472c51ad45b81ebc48706a9a1fe72dd3396322acd0c053b (as mentioned in Ultralytics v8.3.0 release assets
YOLOv11_Nano_Face.onnx	1d09cb0f31d46700a3e80838623aeafa125f2cd0d1c9a12f0b0853bf64b0a83d
YOLOv11_Small.onnx	8bfa953dbe93bef33b09be00da01cb4727f25416cb5d2fdb2a9f1083283b8aaa
YOLOv11_Small_Face.onnx	213333bbecd049c8e1d8bc63b4799df08d2ec52c8f8a6737b37b75ba839d7c03

---

Backend Foundation

• Remove APScheduler from requirements.txt; add psutil and httpx
• app/models/model_registry.py: ModelSpec TypedDict + MODEL_REGISTRY dict
• app/models/model_registry.json: committed registry export for offline bundling
• app/utils/model_downloader.py: ensure_model() with streaming download, SHA-256 verification, and retry (with Range-header resumability).
• YOLO.py: replace eager InferenceSession in __init__ with lazy get_session()
• FaceNet.py: same lazy pattern; on-demand only, not included in first-run tier download

---

Backend API

• app/utils/hardware_detect.py: detect_hardware_tier() using psutil + ort.get_device()
• app/routes/models.py: models_router with:
  • GET /models/status
  • GET /models/recommended-tier
  • POST /models/download/{model_key}: returns task_id
  • GET /models/download/{task_id}/progress: SSE stream
  • DELETE /models/{model_key}
  • POST /models/setup: accepts tier, triggers download of that tier's two YOLO models
• Register models_router in main.py

---

Build and CI

• backend/PictoPy.spec: excludes ONNX_Exports/, bundles model_registry.json
• build-and-release.yml: swap inline pyinstaller command for spec file; remove robocopy/cp model-copy steps
• Verify local build: ONNX_Exports/ absent, model_registry.json present at correct path
• (Optional) Configure CI matrix to output a "nano-bundled" offline installer variant

---

Frontend

• First-run tier selector modal: reads recommended tier, streams SSE progress, dismisses on verified download
• Model manager panel in settings: per-model download/delete with real-time progress bar
• Delete guard: warn if active tier's model is being removed