FashionLens Overview

FashionLens is a lightweight Web application based on multimodal large language models, designed to help fashion designers automatically extract structured metadata from clothing images. It enables intelligent classification, dynamic filtering, and quick retrieval of visual inspiration assets, addressing the pain points of inefficient organization and reuse of massive image collections.

Key Keywords: AI, fashion, multimodal, Gemini, LangGraph, image classification, metadata extraction, fashion design, Pydantic, Flask

For fashion designers, inspiration often comes from casual photos of unique clothing items, fabric textures, or outfit combinations. However, organizing, retrieving, and reusing thousands of inspiration images in phone albums has become a challenge.

Traditional folder classification fails with massive images, and manual tagging is extremely time-consuming. FashionLens was born to automate metadata extraction via AI, allowing designers to focus on creation.

FashionLens's core capability lies in its smart classification system. By calling Google Gemini 2.0 Vision, it automatically extracts 6 key attributes from uploaded images:

• Garment Type: Dress, coat, pants, or accessories
• Style: Retro, minimalist, street, etc.
• Material: Fabric type
• Context & Setting: Suitable occasions and cultural backgrounds
• Color Palette: Main color combinations
• Designer: Brand or designer information

This turns unordered images into a structured retrievable database.

FashionLens uses LangGraph to encapsulate Gemini Vision calls instead of direct LLM calls. This enforces step-based agent flow, enabling future extensions like adding a "Reflection Node" (反思节点) to correct hallucinations or routing nodes for specialized sub-models.

Pydantic models (e.g., GarmentClassification) strictly constrain LLM output formats, ensuring fields like color_palette return lists instead of CSV strings to avoid frontend type errors.

Storage Strategy:

• Images are stored as Blobs in db/images/YYYY-MM-DD/ by date.
• Structured metadata is maintained in images_info.json.
• master_attributes.json cache is generated for UI dropdown menus.

Frontend: Uses native HTML5 (accordion components for dynamic filtering) and client-side search (fast for <10k images). Supports manual custom tags and notes as AI supplements.

Accuracy Tests:

• Garment type: 100%
• Material:100%
• Cultural background:100%
• Occasion:100%

Limitations: Material prediction may be vague without macro texture details. Improvement: Inject style dictionaries into system prompts.

Engineering Tests: Uses pytest for unit (regex markdown stripper), integration (_update_master_attributes), and E2E (upload-classify-filter workflow with mocked Gemini API) tests.

Current Limitations:

1. Client-side search bottleneck for >10k images.
2. Sync upload blocks UI (Flask waits for LLM processing).
3. File system concurrency risks in multi-user scenarios.

Future Plans:

• Migrate search to Elasticsearch/PostgreSQL.
• Use Celery for background tasks.
• Improve storage for concurrency.

FashionLens is a microcosm of AI penetration into vertical industries. It focuses on solving a specific pain point instead of a universal platform. Its "small and beautiful" philosophy, combined with LLM capabilities, provides a solid foundation for secondary development.