Design image and multimodal generation systems

Design an image generation system end to end: cover data collection and curation (sources, licensing, deduplication, filtering, captioning), model architecture choices (e.g., diffusion vs. autoregressive; conditioning and resolution scaling), training objectives and losses, compute/throughput planning, safety and content filtering, evaluation metrics (quality, diversity, bias), inference optimization and deployment (caching, batching, quantization, distillation), cost controls, and monitoring. Then extend the design to a multimodal text-and-image generation system that can accept and produce both modalities. Discuss multimodal data collection and alignment, architectures for cross-modal fusion, training strategies (pretraining, instruction tuning, RLHF/RLAIF), knowledge updating, retrieval augmentation, and product constraints (latency targets, guardrails, feedback loops). Be prepared to walk through a specific recent paper relevant to your design: explain its key idea, experimental setup, metrics, trade-offs, and how you would adapt or productionize it.