A one-person army against model format hell

What it does PINTO_model_zoo is a massive, meticulously maintained collection of pre-converted deep learning models—primarily for computer vision tasks like classification, detection, face recognition, and ReID. Each model ships in multiple formats: TensorFlow Lite, ONNX, OpenVINO, CoreML, TFJS, TF-TRT, EdgeTPU, and plain TensorFlow. The maintainer, PINTO, does this as a hobby, which is either inspiring or slightly alarming.

The interesting bit The real value isn’t the models themselves—most come from other repositories—but the conversion matrix. Need EfficientNet in INT8 for EdgeTPU and FP16 for CoreML? It’s already done, with per-model license files preserved. The README also links to a sprawling ecosystem of custom conversion tools (onnx2tf, tflite2tensorflow, sne4onnx, etc.) that solve specific, painful edge cases like NCHW-to-NHWC transpose explosions or ONNX files exceeding protobuf’s 2GB limit.

Key highlights

• 400+ models across classification, object detection, pose estimation, depth prediction, super-resolution, and more
• Quantization variants: FP32, FP16, INT8, weight quantization, dynamic range quantization, EdgeTPU
• Explicit license tracking per model folder—no mystery meat licensing
• Companion tools for custom ops, model surgery, and format conversion
• DOI-backed (Zenodo) for citation stability

Caveats

• No sample inference code provided; maintainer explicitly skipped this to save time and welcomes PRs
• Some historical notes suggest specific MobileNetV3 and full-integer quantization variants had issues (unclear if resolved)
• Conversion quality varies by model; check the per-folder remarks column for known quirks

Verdict Grab this if you’re deploying vision models to resource-constrained or heterogeneous targets and need format flexibility without wrestling with converters yourself. Skip it if you need turnkey inference examples, training pipelines, or guarantees about model accuracy post-conversion.