dyslexi/Music_covers

Music Cover CLIP Embeddings

CLIP ViT-L/14 (openai/clip-vit-large-patch14) embeddings for ~3.5M music album covers.

Each row contains a concatenated, L2-normalized embedding combining the cover image and the text "{title} by: {artist}":

1. Image is encoded with CLIP's image tower → 768-dim, L2-normalized.
2. Text is encoded with CLIP's text tower → 768-dim, L2-normalized.
3. The two are concatenated → 1536-dim, then L2-normalized again.

Because the final vector is unit-normalized, cosine similarity equals the dot product.

Schema

column	type	notes
id	string	MusicBrainz release-group UUID
title	string	Album title
artist	string	Artist name
embedding	fixed-size list	length 1536

Usage

from datasets import load_dataset
import numpy as np

ds = load_dataset("dyslexi/Music_covers", split="train")
emb = np.asarray(ds[0]["embedding"], dtype=np.float32)  # shape (1536,)

# nearest neighbours of row 0 by cosine similarity
all_emb = np.asarray(ds["embedding"], dtype=np.float32)
scores = all_emb @ emb
top = np.argsort(-scores)[:5]
for i in top:
    print(ds[int(i)]["title"], "—", ds[int(i)]["artist"])