About Decomfy

1. The Reverse Engineering Imperative

Every day, thousands of groundbreaking generative artworks are published to community platforms like Civitai, X, and Reddit. Creators admire the results, download the raw PNG images, drop them into ComfyUI, and are immediately blocked by an avalanche of red error boxes:

Missing custom nodes, undocumented LoRA slider weights, incompatible VAE latents, and vague prompts turn reproduction into hours of painful dependency archaeology. Worse, most published tutorials assume developers have infinite enterprise A100 or H100 cloud VRAM, leaving consumer GPU users stranded with CUDA Out-of-Memory crashes.

Decomfy was built to eliminate this friction entirely through deterministic reverse engineering.

2. The Decompiler Engine & Capabilities

Decomfy decomposes complex generative pipelines into clean, transparent, reproducible engineering assets:

• In-Browser PNG Metadata Decompiler: We extract hidden tEXt: workflow and tEXt: prompt chunks directly from image binary streams in your browser RAM without uploading a single byte to our servers.
• Missing Node Dependency Resolver: Drop any Civitai PNG or canvas JSON into our Missing Node Resolver. Our client-side engine matches missing class names against verified Git repositories and generates 1-click install scripts.
• Interactive Execution DAG Visualizer: Every blueprint features an interactive SVG execution DAG with live pan, zoom, wire tracing, and category filtering (Models, CLIP, Latent, VAE, Image).
• AST Node Graph Diffing: Compare two workflows or prompt iterations side-by-side with our structural AST diff engine to pinpoint parameter changes, widget values, and adapter chains.
• PNG Parameter Injector: Synthesize new ComfyUI-ready PNG files directly in-browser using our PNG Injector, complete with valid 32-bit CRC calculation and embedded prompt DAGs.
• Zero-to-Hero Bootstrapper: Need a clean ComfyUI environment from scratch? Our Bootstrapper converts complex blueprints into automated Bash or PowerShell setup scripts.

3. Consumer GPU Calibration (RTX 4070 Baseline)

Unlike theoretical research papers or corporate demos, every blueprint on Decomfy is benchmarked on physical consumer hardware:

• Primary Calibration GPU: NVIDIA GeForce RTX 4070 (12GB GDDR6X VRAM).
• Host System: Linux x86_64, PyTorch 2.6+cu124, ComfyUI v0.3.x.
• Documented Metrics: Minimum VRAM budget, recommended VRAM budget, peak memory allocation in GB, generation latency per frame in seconds, and exact sampler scheduler combinations.

4. 2026 Technical Roadmap

Decomfy is engineered with an aggressive, multi-phase roadmap to modernize ComfyUI engineering:

• Phase 1: Ingestion & PNG Decompiler (Complete): Client-side PNG chunk extraction, AutoV2/SHA256 hash resolver, desktop bootstrapper, and 8-dimension shell injection regression suite.
• Phase 2: Visual DAG Topology & AST Diff Engine (Current Focus): Interactive SVG execution graphs, dual-workflow AST diffing, prompt token heatmaps, spatial keyboard navigation, and 2026 GEO discovery suite (llms.txt, per-page markdown endpoints).
• Phase 3: Headless Execution & Local VRAM Telemetry (Upcoming): decomfy-cli terminal runner, real-time PyTorch VRAM allocation profiler tracing memory peaks per node pass, and a subgraph flattener converting custom nodes into vanilla ComfyUI core nodes.
• Phase 4: Decentralized Blueprint Registry & Live Radar (Future): Git-driven community workflow submissions without database lock-in, and real-time DiT model velocity tracking across Civitai and HuggingFace.

5. Core Architectural Principles

Decomfy operates under three immutable engineering invariants:

1. 100% Client-Side Privacy: We never collect your private prompts or dropped workflows. All binary parsing and JSON manipulation happen client-side in browser RAM. Zero bytes are uploaded to our servers.
2. Modern Minimalist Standard: No bloated ads, no flashing popups, no muddy gradients, and no raw emojis in UI buttons. We adhere strictly to modern vector iconography (Lucide standard) and clean Linear/Apple aesthetics.
3. White-Weapon BYOP Freedom: We publish 100% SFW educational benchmarks and architecture recipes, while giving users the exact knowledge needed to run unrestricted models on their own private hardware.

6. Get in Touch & Maintainer

Have a workflow you would like reverse-engineered? Found a missing custom node that our resolver did not recognize? Reach out directly:

• Author & Maintainer: Hoang Yell
• Email: hoangyell@gmail.com
• GitHub: github.com/HoangYell
• Domain: https://decomfy.com
• Civitai Showcase: civitai.com/user/hoangyell_com