Style embedder: initial dataset 39 experiment
Two SigLIP2 so400m NaFlex style encoders trained with artist-supervised contrastive learning. baseline is the recommended checkpoint for artist similarity. hsic adds a frozen tagger HSIC term during training. Both were selected at step 23.
| Variant | Held-out artist-pair AUC | Same-artist pairs | Different-artist pairs |
|---|---|---|---|
baseline |
0.9077 | 298 | 119,507 |
hsic |
0.8968 | 298 | 119,507 |
The evaluation used 490 images from 376 artists excluded from training. Checkpoint selection used this validation set, so these figures are not an independent external test. The training subset contained 1,477 images from 209 artists. No source images or artist identities are included here.
Each variant has:
best_style_encoder.safetensors: complete encoder with its LoRA weights merged. Use this for inference.best_metadata.json: architecture and preprocessing required by the loader.best_style_lora.safetensors: trainable parameters, including LoRA, pooling, and projector weights. This is not a standalone image-generation LoRA or an exact-resume checkpoint.
For inference in SushiUI, pass a variant's best_style_encoder.safetensors to core.style_encoder.style_artifacts.load_style_encoder; keep best_metadata.json beside it. The complete encoder needs no base-model download. The LoRA file requires the same base SigLIP2 weights and StyleEncoder architecture if loaded separately.