The Context
The Innovation Pitch documents what happened when a routine assignment -- "learn ComfyUI and come back and share what you found" -- turned into a week-long deep dive that produced a trained AI model, a ComfyUI plugin, a 7-day course draft, a documentary, and physical art for a sold-out charity gala. The presentation was designed for leadership at a creative agency, making the case for integrating AI image generation tools into production workflows.
The Learning Arc
Days 1-2 covered ground-up installation and fundamentals: source install from Git (Homebrew, Python 3.12, virtual environments), understanding the model ecosystem (checkpoints, LoRAs, VAEs, CLIP encoders), ComfyUI Manager for custom nodes, and building txt2img/img2img workflows from scratch. Day 3 is where it escalated: training a custom LoRA model on 13 images of the Savoy Ballroom sign across three training runs at approximately two dollars per run on fal.ai cloud GPUs, with attention to epochs, steps, dataset quality control, watermark removal, and caption consistency.
The Plugin
The technical centerpiece is a ComfyUI sidebar panel that talks to Claude, builds workflows directly on the canvas from natural language, and learns from its own mistakes. The architecture: JavaScript frontend (chat panel, error capture, workflow loading), Python backend (Claude API integration, prompt builder, model downloads), and SQLite knowledge base with FTS5 full-text search for auto-learning workflow patterns.
The plugin intercepts ComfyUI errors through three layers -- WebSocket, fetch, and DOM observation -- and suggests fixes. It stores error-to-fix patterns so the next time the same error hits, it already knows the answer. About 3,200 lines of code across 12 source files, with 7 database tables (workflow_patterns, lessons, node_specs, chat_history, error_log, knowledge_fts, github_sync) and 760+ searchable knowledge entries.
The Broader Output
The presentation maps all the outputs: a working ComfyUI knowledge base, a course draft shared with the agency's Learning Director, the ComfyUI plugin prototype, workflow templates, and the Savoy Ballroom gala work (detailed in the separate Savoy Ballroom Display project). It also references existing internal tools already in use: an animation automation script saving approximately 45 minutes per use, a cross-pollination script for the dev team saving approximately 90 minutes per session, and the Mo-BrAInz AI assistant panel for After Effects.
The Pitch
The presentation closes with a proposal for a Unified Task Intelligence System -- a locally-hosted tool (what became SpaceTable) that captures tasks from Teams, Airtable, Frame.io, and manual input using a local AI model (Ollama), tracks everything in a local SQLite database, and generates reports automatically. The pitch emphasizes that the local AI model runs entirely on-machine with no cloud dependency, zero subscription cost, and full IT visibility.
Tech Stack
PowerPoint (presentation), ComfyUI (Python), Claude API, SQLite with FTS5, JavaScript/Python plugin architecture, fal.ai (cloud GPU training).