VML (Visual ML) β v2.0.0
VML (Visual ML) is a 100% offline ConvNeXt-Tiny multi-head ONNX model for Mobile Legends: Bang Bang (MLBB) visual asset recognition, distilled from 768-d Vision Teacher embeddings (vit_base_patch16_siglip_224).
Capabilities
- 4,677 standardized PNG training assets across 134 heroes (icons, portraits, skin portraits, skills), items, battle spells, emblems, talents, and UI elements.
- 282 coarse classes (
<Hero>,Item,Spell,Emblem,Talent,Skill | <Hero>,UI). - 2,794 fine-grained classes (
<Hero> | skin_XX,<Hero> | icon,<Hero> | portrait,Item | <Name>,Spell | <Name>,Emblem | <Name>,Talent | <Name>).
Benchmark & Tournament Stress-Test (v2.0.0)
Evaluated on 200 synthetic tournament broadcast samples (dark bottom gradient, IGN text overlay, corner spell occlusion, circle/rounded crop, OBS Gaussian blur, JPEG compression):
- Coarse Top-1: 99.0% (
0.99) - Fine (Hero + Skin / Item / Spell) Top-1: 93.0% (
0.93) - 768-d DB Retrieval Top-1: 99.5% (
0.995) - 768-d DB Retrieval Top-5: 100.0% (
1.0)
Files
hero_skin_student.onnx+hero_skin_student.onnx.dataβ ONNX model (pixel_values[B, 3, 224, 224]->embedding[B, 768],coarse_logits[B, 282],fine_logits[B, 2794]).hero_skin_db.npzβ 768-d normalized reference embeddings(4677, 768).hero_skin_meta.jsonβ Metadata mapping each DB row tocoarse,fine, andpath.hero_skin_coarse_labels.json&hero_skin_fine_labels.jsonβ Label indices.training_summary.jsonβ Training & stress-test metrics.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support