Research map / Multi-agent systems and games / LLM agents
StyleTailor: Towards Personalized Fashion Styling via Hierarchical Negative Feedback
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes StyleTailor, a training-free multi-agent framework for personalized fashion styling that combines garment retrieval, outfit generation and FLUX.1 Kontext virtual try-on, refined by hierarchical VLM negative feedback at item, outfit and try-on levels.
Problem
Personalized fashion styling (garment design, shopping recommendation, virtual try-on, evaluation) has no unified agent framework, and existing agentic refinement relies on random or repetitive output selection that is costly and yields limited gains.
Approach
A Designer agent (sequential expert VLMs plus Google Custom Search retrieval) and a Consultant agent (FLUX.1 Kontext try-on) refine outputs through closed-loop negative feedback: failed retrievals and try-ons are converted into negative prompts or passed as negative examples to the next expert, with a VLM-based evaluation suite.
Key results
- Full StyleTailor reaches style consistency 0.906 vs 0.650 for the baseline without negative feedback, and face similarity 0.544 vs 0.362.
- VLM Artist score improves from 7.35 (baseline) to 8.60 for the full system.
- Removing try-on-level feedback drops visual quality from 0.764 to 0.708 and face similarity from 0.544 to 0.351.
Surprising findings
- Removing item-level negative feedback slightly increases visual quality (0.767 vs 0.764 for the full system), while lowering style consistency sharply.
Limitations
- Performance is bounded by the capabilities of the pre-defined VLMs in the system.
- Evaluation uses a small dataset of 64 samples and a single VLM judge (qwen-vl-max) for the artist score.
Experimental setup
Datasets: LookBook
Benchmarks: LookBook (64-sample evaluation set)
Metrics: style consistency (VQAScore) visual quality (IQAScore) face similarity (InsightFace cosine) VLM artist score
Baselines: Baseline without negative feedback
Builds on: FLUX.1 Kontext VQAScore CLIPScore InsightFace
Closest papers by meaning
- PICTURE: PhotorealistIC virtual Try-on from UnconstRained dEsigns (CVPR 2024 · similarity 0.85)
- HieraFashDiff: Hierarchical Fashion Design with Multi-stage Diffusion Models (AAAI 2025 · similarity 0.85)
- FashionTailor: Controllable Clothing Editing for Human Images with Appearance Preserving (AAAI 2025 · similarity 0.84)
- Towards Unified Multi-Modal Personalization: Large Vision-Language Models for Generative Recommendation and Beyond (ICLR 2024 · similarity 0.83)
- Correpondence Alignment For Improved Virtual Try-On (NeurIPS 2026 · similarity 0.83)
- IMAGDressing-v1: Customizable Virtual Dressing (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.