crisng95/flowboard

What it solves

Flowboard is designed for repetitive e-commerce video production where maintaining consistency across characters, products, and scenes is difficult. It replaces the manual process of re-uploading references and re-typing prompts for every variant with a graph-based workflow, ensuring that a single source of truth for assets (like a specific model or garment) persists across an entire campaign.

How it works

The project uses an infinite-canvas workspace where assets are treated as nodes in a directed graph:

• Reference Nodes: Users set up "Character" and "Visual asset" nodes once. These act as anchors; a vision-capable LLM describes these assets to create an aiBrief that is passed downstream.
• Composition Nodes: By connecting reference nodes to an "Image" node, the system automatically synthesizes fashion-editorial prompts using a connected LLM CLI (e.g., Claude Code). It generates multiple pose-distinct variants in one batch.
• Video Nodes: These take an image as input and use Google's Veo 3.1 i2v to generate video. The system uses "motion vocabulary" tailored to the detected scene (e.g., studio vs. street) to ensure natural movement.
• Proxy Architecture: Because it targets Google Flow, it uses a Chrome extension to proxy authenticated requests and reCAPTCHA tokens from the user's browser to a local FastAPI agent.

Who it’s for

It is built for e-commerce creative producers and AI artists who use Google Flow (Pro/Ultra plans) to create consistent product and character-driven video content.

Highlights

• Graph-Based Workflow: Every asset is a reusable node, and every edge represents a real data dependency.
• Environment-Aware Prompting: Automatically switches motion vocabulary based on whether the scene is a studio, street, café, or nature setting.
• Local-First Privacy: All board state and media are stored locally via SQLite and a local storage folder.
• CLI-Based LLM Integration: Shells out to existing LLM CLIs (Claude Code, Gemini CLI, OpenAI Codex) to avoid direct cloud API costs.
• Batch i2v Generation: Can spawn multiple video variants from a single click if the upstream image has multiple variants.