HIRING: AI Inference Engineer — Real-Time Diffusion Pipeline

SmokeScreen AI is a real-time AI media engine for content creators and streamers. We are building a live full-frame identity transfer pipeline, think: your movements, someone else’s full appearance, rendered in real time. The core pipeline is already architected. We need an engineer who can implement, optimize, and benchmark it on GPU infrastructure. THE WORK You will build and optimize a real-time inference pipeline consisting of: • Stable Diffusion 1.5 backbone • ControlNet (OpenPose + Canny) for structural conditioning • IP-Adapter FaceID for identity injection • LCM scheduler (2-step inference) • FastAPI WebSocket server for live frame streaming • Deployed on RunPod GPU infrastructure (A100) The adapter architecture is already designed. Your job is implementation, optimization, and getting us to the highest possible FPS on an A100. STRONG PREFERENCE ━━━━━━━━━━━━━━━━━━━━━━━━━ ⭐ StreamDiffusion experience ⭐ LCM / distilled diffusion models ⭐ InsightFace / ArcFace embeddings ⭐ TensorRT optimization ⭐ Prior real-time video AI project ━━━━━━━━━━━━━━━━━━━━━━━━━ HOW TO APPLY ━━━━━━━━━━━━━━━━━━━━━━━━━ Do NOT send a generic proposal. To be considered, answer these three questions in your application: 1. Have you worked with IP-Adapter FaceID before? If yes, describe what you built. 2. What is the realistic FPS you can achieve with SD1.5 + ControlNet + LCM on an A100 80GB at 512×512, and how would you push it higher? 3. What is your availability this week? Applications without answers to these questions will not be reviewed.