ZooClaw-FashionSigLIP2: Distilled Fine-tuning for Robust Fashion Retrieval

A team of researchers has released ZooClaw-FashionSigLIP2, a vision-language model specialized for fashion retrieval that outperforms competing approaches, including larger models, on every benchmark in its evaluation suite [1]. The model, detailed in a paper submitted to the arXiv preprint repository on 26 Jun 2026, adapts the SigLIP2-base architecture using a training recipe of full fine-tuning with knowledge distillation on curated in-domain data, followed by weight interpolation with the base model [1]. The authors report that this approach surpasses parameter-efficient methods such as LoRA, larger backbones with up to 1B parameters, and models trained on external datasets [1]. The work addresses a core tension in adapting foundation models: improving performance on a specialized target distribution without sacrificing the broad generalization of the original encoder, a challenge the authors describe as particularly stringent in fashion retrieval [1]. Alongside the model, the researchers introduce ZooClaw-Fashion, a new benchmark for fashion retrieval, and present a systematic quality analysis of existing benchmarks that identifies and mitigates structural biases in their public ground truth [1]. The model weights and all evaluation artifacts have been open-sourced [1]. The paper appears on arXiv, an open-access repository that hosts electronic preprints across disciplines including computer science and has grown to receive approximately 24,000 submissions per month as of late 2024 [6]. The repository is not peer-reviewed; papers are approved for posting after moderation [6]. The release follows a pattern of community-driven tool development around the preprint platform. Since 2020, arXiv has operated arXivLabs, a framework that enables third-party collaborators to build experimental features on the article record page, such as citation explorers and code finders, under guidelines that require adherence to openness and user data privacy [5]. The ZooClaw-FashionSigLIP2 abstract page displays several such Labs integrations, including Bibliographic Explorer and Connected Papers [1].