Flippd Pinball Vision identifies the pinball machine shown in an image. Release 1.0.0 was published on 2026-07-31. The private training corpus is not distributed.
The repository contains two supported ONNX artifacts. The root FP32 graph is the Python/server/CPU model. The browser FP16 graph is the browser/mobile model; it is not a CPU/server replacement, because source ONNX Runtime CPU measurements used materially more memory.
Both graphs use the dinov3-vits16 encoder, ONNX opset 18, label schema 2, and the same ordered vocabularies. Both passed the full ONNX checker. The browser graph uses FP16 weights and general compute while retaining exactly 24 attention-score MatMul/Softmax nodes in FP32 to prevent observed FP16 overflow; its inputs and outputs are FP32. It prefers WebGPU and requires a WASM fallback.
Download and installation
Access is gated. Request access on Hugging Face, accept the license, and authenticate before downloading:
Python >=3.12 is supported. Install the standalone FP32 example dependencies:
python -m pip install -r requirements.txt
For CUDA, replace onnxruntime with onnxruntime-gpu; do not install both in the same environment.
FP32 Python usage
CLI
python inference.py ./photo.jpg --top-k 5 --device cpu
Optional controls include --threads N and --device cuda or --device cuda:N. The command prints indented JSON and exits nonzero on image decoding, checksum, provider, validation, or inference failure.
PinballClassifier defaults to the model and metadata beside inference.py. It verifies the metadata contract and model SHA-256 before creating an ONNX Runtime session.
Browser usage
The browser example must be served over static HTTP; opening index.html as a local file is not supported. From the downloaded repository root:
The page fetches the model and metadata only from the downloaded repository on the same origin. No Hugging Face token enters browser code, and the page makes no Hub API or remote model request. Network access is required only to load the pinned ONNX Runtime Web 1.22.0 JavaScript/WASM runtime assets from jsDelivr; model access and inference remain local.
Inputs and preprocessing
The Python path decodes with Pillow and converts to RGB. It resizes the short edge to 256 while preserving aspect ratio, using integer-truncated dimensions and bicubic interpolation, then center-crops 256 × 256 with offsets round((dimension - 256) / 2.0). It converts HWC to contiguous CHW float32, divides by 255, and normalizes channels with:
mean: [0.485, 0.456, 0.406]
standard deviation: [0.229, 0.224, 0.225]
The browser follows the same resize, crop, RGB, layout, scaling, and normalization contract using a canvas with high-quality smoothing. Browser canvas resampling is not byte-identical to Pillow, so preprocessing parity is not claimed. Neither implementation applies EXIF orientation normalization. Pillow-supported image codecs vary by installation.
Outputs and labels
Each inference returns three independent ranked heads in this order:
exact_logits: 1,499 exact classes (G<group>-M<machine>[-A<alias>] or __unknown__)
The identifiers are canonical Open Pinball Database (OPDB) identifiers. Group, machine, and alias components retain OPDB's G, M, and A prefixes and are combined cumulatively as shown above; __unknown__ is the sole non-OPDB sentinel.
The ordered vocabularies in each artifact's onnx-metadata.json are the authoritative class maps. Their canonical vocabulary SHA-256 is 7f0de4a0cc94845d5e6f429ca9c6eac81dbef4e7cdfe5f008267ce19d78c1cc1. Logits are already temperature-scaled; clients do not apply the metadata temperatures again. The examples compute a stable softmax independently for each head, sort descending, and return one to twenty { "id", "confidence" } candidates with confidence rounded to six decimals. __unknown__ is an ordinary exact class, not a general detector for inappropriate or out-of-distribution images.
The heads are independent rankings and provide no acceptance decision. Applications should expose confidence and use appropriate human review rather than silently treating a candidate as certain.
Evaluation
These are source-observed export results, not measurements made with browser canvas preprocessing:
Evaluation
Result
Sealed 63,038-image exact top-1, FP32
91.0038%
Sealed 63,038-image exact top-1, FP16
90.9927%
Exact top-5 after FP16 conversion
Unchanged
Top-1 parity across group, machine, and exact heads on 200 deterministic held-out images
100%
Top-1 parity does not imply equal logits or confidence values.
Limitations
Predictions may be wrong for visually ambiguous, cropped, low-quality, multiple-machine, unusual-viewpoint, remake, alternate-art, or out-of-distribution images. Detail images, occlusion, and poor lighting may remove the evidence needed for identification. __unknown__ does not reliably identify all unsuitable inputs. Callers should show confidence, preserve alternatives where useful, and require human review when an error matters.
The model identifies visual evidence associated with a machine. It does not establish ownership, location, authenticity, edition provenance, market value, or whether every visible component belongs to the predicted title. Do not use it for unlawful activity, biometric identification, surveillance, attempts to recover private training data, or as a substitute for meaningful human review in consequential decisions.
License, support, and attribution
The model materials use the Flippd Community Model License 1.0, which is not an open-source license. Personal, educational, research, and genuinely non-commercial community use is permitted under its terms. Commercial use—including paid products, business operations, advertising, sponsorships, consulting, and other commercial benefit—requires a separate written license; contact admin@flippd.gg.
A free community project may receive optional tips or Patreon-style support without becoming commercial only while every license condition remains satisfied, including materially equal service for contributors and non-contributors, no purchased benefits, no advertising/sponsorship/affiliate or required-payment revenue, and no more than US $5,000 gross qualifying support in any rolling twelve-month period. See LICENSE.md for the complete terms.
Public projects must display:
Powered by Flippd Pinball Vision.
An About, Credits, Acknowledgments, documentation, or README location must also state:
This project uses the Flippd Machine Classifier, developed by the Pindigo/Flippd team under the Flippd Community Model License 1.0.
Privacy
Running downloaded weights locally does not, by itself, send images, predictions, or other model inputs to Flippd. Hugging Face provides Flippd with the username, email address, access-form answers, and related access information described in PRIVACY.md. Flippd does not use gate-request information for unrelated marketing or sell it. Flippd's general privacy policy is at https://flippd.gg/privacy-policy. Privacy requests may be sent to admin@flippd.gg; do not send passwords or access tokens.
Citation
bibtex
1@misc{flippd_machine_classifier_2026,
2 author = {Flippd, LLC},
3 title = {Flippd Pinball Vision — Machine Classifier},
4 year = {2026},
5 version = {1.0.0},
6 howpublished = {Hugging Face model repository},
7 url = {https://huggingface.co/Flippd/pinball-classifier}
8}
