Compare PyTorch vs ONNX inference on a single image. Toggle variants to see how predictions and speed differ across backends without re-running the PyTorch model.

Batch-test multiple images across ONNX variants. Measures per-variant inference speed (avg / max ms) and prediction agreement to help you choose the right model for deployment.

Model: buildborderless/CommunityForensics-DeepfakeDet-ViT — ViT-Small classifier for deepfake image detection, trained on 2.7M samples across 4,803 generators. Outputs real/fake with sigmoid confidence.

Compare tab

Upload a single image to see inference results side-by-side from the PyTorch model and your selected ONNX variants.

• PyTorch runs the full-precision model on CUDA (or CPU if no GPU).
• ONNX runs pre-exported variants optimized for different deployment scenarios.
• Toggling ONNX variants re-uses the cached PyTorch result — only ONNX re-runs.
• Click All / None to quickly select or clear variants.
• Expand Raw JSON to see the exact numeric output for scripting.

Benchmark tab

Upload multiple images and run batch inference across selected ONNX variants.

• Each image runs through every selected variant: 5 warmup runs + 10 timed runs.
• Predictions table shows per-image, per-variant verdicts with real/fake probabilities.
• Avg / Max ms footer row shows mean inference time and worst-case latency.
• Uploaded images appear in the thumbnail gallery above the upload control.
• Use Select All / Deselect All and the Run / Clear buttons to iterate quickly.

ONNX variant guide

Five pre-exported ONNX models with different size/speed trade-offs. All use the corrected config (single-class sigmoid output).

Which variant should I use?

Running locally

pip install gradio torch transformers onnxruntime huggingface_hub
python app.py