SentenceTransformer Target Encoding

Overview

In image-to-prompt competitions, the target is not raw text but its SentenceTransformer embedding. Models predict embedding vectors and are scored by cosine similarity against the ground-truth prompt embeddings. Understanding this encoding step is essential: predictions must live in the same embedding space as the evaluation targets.

Quick Start

from sentence_transformers import SentenceTransformer
import numpy as np

st_model = SentenceTransformer("all-MiniLM-L6-v2")

prompts = ["a painting of a sunset over mountains", "cyberpunk city at night"]
embeddings = st_model.encode(prompts, normalize_embeddings=True)
# embeddings.shape: (2, 384)

# Cosine similarity between two prompt embeddings
similarity = np.dot(embeddings[0], embeddings[1])

Workflow

1. Load a SentenceTransformer model matching the competition's evaluation model
2. Encode all ground-truth prompts into dense vectors (the target)
3. Any prediction pipeline must output vectors in this same space
4. Score predictions via cosine similarity: np.dot(pred, target) / (norm(pred) * norm(target))
5. For submission, flatten the embedding matrix row-wise into a single column

Key Decisions

• Model choice: must match evaluation exactly — all-MiniLM-L6-v2 (384-dim) is common; check competition description
• Normalization: enable normalize_embeddings=True to simplify cosine similarity to dot product
• Batch encoding: use batch_size=32 for large datasets to avoid OOM
• Style keywords: appending style tokens ("fine details, masterpiece") to captions before encoding can bias embeddings toward the target distribution
• vs. CLIP text embeddings: SentenceTransformer captures semantic meaning; CLIP captures visual-textual alignment — different spaces

References

• BLIP+CLIP | CLIP Interrogator
• Calculating Stable Diffusion Prompt Embeddings