Skip to content

README.md

Latest commit

 

History

History
109 lines (89 loc) · 3.69 KB

README.md

File metadata and controls

109 lines (89 loc) · 3.69 KB

Vision chain

Project Overview

Overview

Python, FastAPI, Uvicorn, LangChain, LangGraph, Google Gen AI, React, TypeScript, Vite

Sandbox project built to experiment with multimodal AI pipelines and stateful orchestration.

The goal is to take a user-uploaded portrait, analyze it, and dynamically transform it using a graph-based agent workflow

LangSmith

The Stack & Why

  • LangGraph & LangChain: Handles the agent workflow, state management, and routing logic between execution nodes
  • Google Gen AI (gemini-2.5-flash-image): Used both as a vision tool to analyze the image and as a generative engine for the final image-to-image modification.
  • LangSmith: Hooked up for quick debugging, viewing execution traces, and tracking agent states in real-time.

The Flow

  1. Analyze: The system inspects the input image to detect visible features (skin, jawline, hair)
  2. Route: It identifies the subject's gender to branch the state into a specific processing path
  3. Enhance: It builds a tailored prompt based on the analysis and fires it at the Gemini vision model with customized safety overrides to generate a high-end, edited version of the portrait

Features

Feature Analysis

Graph Nodes & Agents Workflow

  • image_cache: Uploads the image and creates a 5-minute Gemini Context Cache if the file exceeds 1024 tokens to save bandwidth and API costs
  • gender_classification: A zero-temperature vision gatekeeper that detects if a human is present and returns their gender (male, female, or none) in raw JSON
  • feature_analysis: Pydantic validation schemas dynamically based on gender to detect visible face/body parts (eyes, jawline, skin, hair)
  • image_retouch_specifier: Conditional prompt builder. It maps the detected features into pro-grade retouching instructions
  • execute_image_enhancement: The execution node. Fires the final prompt at gemini-2.5-flash-image, bypasses false-positive safety locks, and saves the output image

State Management

The entire workflow relies on a single, stateful AgentState object managed by LangGraph. Each node in the graph reads from this shared context and returns only the fields it intends to update

       [Input State]
             │
             ▼
      (image_cache)
             │
             ▼
  (gender_classification)
             │
             ▼
     (feature_analysis)
             │
             ▼
 (image_retouch_specifier)
             │
             ▼
(execute_image_enhancement)
             │
             ▼
      [Output State]

Configuration

GOOGLE_API_KEY=
LANGCHAIN_TRACING_V2=true
LANGSMITH_ENDPOINT=https://eu.api.smith.langchain.com
LANGCHAIN_API_KEY=lsv2_pt_
LANGSMITH_PROJECT="vision-chain"

Usage

cd backend && uvicorn main:app --reload --reload-dir .
cd frontend && npm run dev

Project Structure

vision-Chain/
├── backend/
│   ├── routers/
│   │   ├── agents/
│   │   │   ├── graph.py
│   │   │   ├── nodes.py
│   │   │   ├── router.py
│   │   │   ├── schemas.py
│   │   │   └── state.py
│   │   └── upload.py
│   └── main.py
├── frontend/
│   ├── src/
│   │   ├── App.tsx
│   │   ├── index.css
│   │   └── main.tsx
│   ├── index.html
│   ├── package.json
│   ├── tsconfig.json
│   └── vite.config.ts
├── static/
├── storage/
├── uploads/
├── .env
└── .gitignore

License

This project is licensed under the MIT License - feel free to do whatever you want with the code