OppaiOracle V1.1 — anime tagger

Multi-label ViT fine-tuned at 448×448 (a fine-tune of the from-scratch V1 320×320 model, on the same cleaned ~5.9M-image corpus, 19,294 tags). Drop in an image to see ranked tag predictions.

Read first: the model card on the model repo documents known noise patterns (color tags, hair-length boundaries, neckwear, missing-tag bias). Predictions are best treated as a fast first pass that a human reviews — not as ground truth.

Image

Threshold

Tags below this score are dropped. Macro PR break-even on val ≈ 0.760.

0.05 0.95

Max tags

5 200

Predictions

Comma-separated tags

Model: Grio43/OppaiOracle · Resolution: 448×448 · Tags: 19,294 · Activation: sigmoid (already applied inside the ONNX graph).

This Space runs the V1.1 (448×448) checkpoint. The V1 (320×320) checkpoint is also in the model repo for users who specifically want the smaller native resolution — match input resolution to the checkpoint you load.