CLIP faces TCG card-scan doubts

CLIP is a solid first-pass retriever, but it is probably too blunt to be the final identifier for TCG cards. The hard part here is fine-grained discrimination, not broad semantic similarity.

• –OpenAI’s own CLIP writeup says it excels at zero-shot generalization but struggles with fine-grained classification and OCR, both core to card lookup.
• –Near-duplicate printings, foils, languages, and set symbols make a pure embedding match fragile.
• –The thread’s suggested alternative, Qwen3-VL-Embedding, is built for multimodal retrieval and reranking, which is a more direct fit.
• –A hybrid pipeline, embeddings for recall plus OCR or reranking for confirmation, will usually beat CLIP alone.