Image Transfer is a local, browser‑based image‑to‑image style transfer app that runs ONNX models via onnxruntime‑web. It supports quantized INT8 models with safe fallbacks to FP32.

Run The App

• Dev server: npm run dev (Next.js 15 + Turbopack)
• Open: http://localhost:3000
• Use: pick a style and an image, adjust Strength slider.

Models & Presets

• Families:
  • agan: AnimeGANv3, dynamic NHWC input (studio look styles: Ghibli, Hayao, Shinkai, Sketch)
  • fns: Fast Neural Style, fixed 224×224 NCHW (Mosaic, Candy, Udnie, Pointilism)
• Paths:
  • Quantized INT8: public/models-int8/*.int8.onnx
  • FP32 originals: public/models-agan/*.onnx (AnimeGAN) and public/models/*.onnx (FNS)

INT8 vs FP32 Fallback

• The status line shows which model is active after load:
  • “Ready: <Style> (INT8)” → running the quantized model.
  • “Ready: <Style> (FP32 fallback)” → quantized model failed to load; using FP32.
• Console diagnostics:
  • Failures summarize as one line prefixed with [ORT-LOAD-FAIL] { ... }.
  • Enable verbose ORT logs only if needed: add ?ort_debug=1 to the URL or run localStorage.setItem('ort_debug','1') and reload.

Quantization (Python)

Quantization scripts live in python/. We use gentle, WASM‑friendly settings and per‑family defaults.

• Environment (suggested):

  • python -m venv .venv && source .venv/bin/activate
  • pip install onnx onnxruntime onnxruntime-tools pillow numpy

• Command (AnimeGAN, higher precision but stable):

  • python python/quantize_models.py --src public/models-agan --dst public/models-int8 --only agan_ghibli.onnx,agan_hayao.onnx,agan_shinkai.onnx --samples 64 --size 192 --percentile 99.99
  • Defaults used per family:
    • agan: QOperator (QLinearConv), Conv only, per‑tensor (per‑channel off), opset 13.
    • fns: QOperator (QLinearConv/QLinearMatMul), Conv + MatMul, per‑channel on, opset 13.

• Command (FNS styles):

  • python python/quantize_models.py --src public/models --dst public/models-int8 --only fns_candy.onnx,fns_mosaic.onnx,fns_udnie.onnx,fns_pointilism.onnx --samples 64 --size 224 --percentile 99.9

• Useful flags:

  • --only filter by basenames/globs (comma‑separated).
  • --samples calibration images to use (static only). More samples → better fidelity, slower.
  • --size calibration size. Larger → better fidelity, slower.
  • --percentile activation percentiles (e.g., 99.9). Higher reduces clipping.
  • --force_per_channel enable per‑channel weights even for AnimeGAN.
  • --qdq emit QDQ instead of QOperator (generally keep QOperator for WASM).

Notes:

• The script pre‑upgrades FP32 models to opset 13 before quantization for cleaner graphs.
• It may emit temporary *.preopset.onnx files in public/models-int8/; these can be deleted.

NPM Script Shortcuts

• npm run quant:agan → AnimeGAN (all): samples=64, size=192, percentile=99.99
• npm run quant:fns → FNS (all): samples=64, size=224, percentile=99.9
• npm run quant:ghibli → Ghibli only (same settings as agan)
• npm run quant:clean → remove any *.preopset.onnx intermediates

onnxruntime‑web Behavior

• Execution provider order: WebGPU (if available) then WASM.
• WASM uses SIMD and a single thread by default (no COOP/COEP needed).
• If an INT8 model fails to load under WASM, the app automatically falls back to the FP32 model and marks the status as “(FP32 fallback)”.

Troubleshooting

• Noisy console: default logs are minimal. Turn on deep logs with ?ort_debug=1 only when requested.
• Share failures by copying the single [ORT-LOAD-FAIL] line from the console.
• If a specific model keeps falling back, try gentler quantization (reduce ops, turn off per‑channel, lower percentile) and re‑quantize that model with --only.