Milo — CCG Card Embedder

MobileViT-XXS backbone trained with ArcFace loss (multitask: illustration_id + set_code) to produce 128-dimensional L2-normalised embeddings of CCG card images for nearest-neighbour retrieval.

Model details

Property	Value
Architecture	MobileViT-XXS + linear projection
Input	448×448 RGB, ImageNet-normalised
Output	128-d L2-normalised embedding vector
Parameters	~1.0M
File size	5.2 MB (fp32 ONNX)
Codename	milo
Version	1.0.0 (epoch 15)
Training labels	illustration_id + set_code (multitask ArcFace)

Usage

The easiest way to use Milo is through the CollectorVision library, which handles corner detection, dewarping, gallery loading, and nearest-neighbour search:

import collector_vision as cvg

cvid = cvg.Identifier(cvg.HFD("HanClinto/milo", "scryfall-mtg"))
result = cvid.identify("photo.jpg")
print(result.ids)         # {"scryfall_id": "..."}
print(result.confidence)  # 0.94

Direct ONNX usage

import onnxruntime as ort
import numpy as np
from PIL import Image

session = ort.InferenceSession("model.onnx")

# Preprocess: resize to 448×448, ImageNet normalise, NCHW float32
img = Image.open("card_crop.jpg").convert("RGB").resize((448, 448))
x = np.array(img, dtype=np.float32) / 255.0
x = (x - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]
x = x.transpose(2, 0, 1)[None]  # (1, 3, 448, 448)

emb = session.run(None, {"pixel_values": x})[0]  # (1, 128) float32, L2-normalised

Cosine similarity between two embeddings is just their dot product (both are unit vectors).

Gallery compatibility

Gallery files built with Milo v1.0.0 use milo1 in their filename. Embeddings from different Milo versions are not compatible — rebuild the gallery when upgrading.

Part of CollectorVision

Used together with HanClinto/cornelius in the CollectorVision inference library.