Show HN: Visual autocomplete for drawings (real-time Human-AI interaction) (github.com)

A new open-source C++ app (AiDrawing) demonstrates real-time, prompt-free “visual autocomplete” for freehand sketches by closing a live loop between user strokes, a vision model, and image-to-image diffusion. Built on the Cinder creative-coding framework, the app captures vector strokes (pressure, smoothing, undo/redo) and sends a raster of the canvas to local Ollama vision models (e.g., LLaVA) for semantic interpretation; those descriptions feed a StreamDiffusion Spout server that performs GPU-accelerated image-to-image diffusion and streams augmented textures back via Spout for onscreen compositing. Communication is asynchronous and multi-threaded (ThreadSafeList) and supports OSC for integration with TouchDesigner/Max, enabling low-latency, continuous human–AI collaboration without text prompts. This is significant for the AI/ML and creative-coding communities because it showcases a practical architecture for live, privacy-preserving generative assistance: local inference (Ollama), zero-copy GPU texture sharing (Spout), and a separate diffusion server let artists iterate interactively with model-driven completions. Key technical notes: Windows is required for Spout; recommended GPU ≥8GB VRAM; project includes an Ollama C++ client (OllamaClientCinder), uses command-pattern undo/redo, and supports switching vision models (llava:7b/13b). The repo and instructions make it easy to reproduce the pipeline, offering a blueprint for integrating multimodal models into real-time creative tools.