Price: 2500000
Number of applications: 1
26.11.26 (inclusive)
Milestone-based fixed contract + performance bonus for achieving benchmark FPS
MVP
ICT tasks
Media sphere
Neurotechnology and artificial Intelligence
Purpose and description of task (project)
Architecting and implementing an end-to-end, high-performance inference pipeline for seamless video face swapping and reenactment on high-resolution streams (1080p / 4K). Project Scope: - Benchmarking and orchestrating state-of-the-art open-source architectures (InsightFace, SimSwap, Ghost, LivePortrait). - Incorporating facial restoration and super-resolution models (CodeFormer, GPEN) to preserve fine texture and eye/mouth clarity. - Implementing temporal consistency filters and optical flow constraints to eradicate inter-frame jittering and blending artifacts. - Delivering a containerized production service with NVIDIA TensorRT/CUDA acceleration exposing REST and gRPC endpoints.
Software/ IS
Out-of-the-box open-source implementations suffer from severe temporal inconsistency (frame flickering), occlusion blending failures, and artifacts at extreme head poses (yaw/pitch > 45°), paired with sub-optimal inference speeds. Conversely, enterprise cloud APIs incur prohibitive per-minute usage fees and introduce data privacy risks. There is a critical need for an in-house, low-latency, temporally stable pipeline that delivers superior identity preservation (Face ID retention) and clean edge blending on standard server-grade GPUs.
1. Sustained inference throughput >= 24 FPS at 1080p on single NVIDIA RTX 4090 / A10G accelerators. 2. Identity preservation score (Face ID retrieval via Cosine Similarity) >= 95%. 3. Complete mitigation of mask-boundary artifacts and inter-frame temporal flickering. 4. Over 70% cost reduction per processed video stream compared to external SaaS APIs. 5. Production-ready microservice equipped with clear OpenAPI / Swagger documentation.
Aidos Kuanyshev
Note
Demonstrated proficiency with PyTorch, TensorRT/ONNX Runtime, and face parsing models is required. Dedicated cloud GPU instances are provided.
Whatsapp
