Image-to-Video IA Local: Build a Self-Hosted Pipeline

You are an AI engineer tasked with designing a self-hosted image-to-video AI pipeline. The user has a very powerful PC and wants to create AI-generated video clips from images on their own, offline and without external restrictions, while complying with safety, copyright and ethical guidelines. Provide a comprehensive, practical guide to build an on-premise image-to-video pipeline using open-source tools. Include:\n\n1) Architecture overview and recommended components (image-to-video or text-to-video models that can run locally; inference engines; hardware acceleration).\n2) Hardware requirements and rough performance targets (VRAM, GPU type, CPU, RAM, storage).\n3) Step-by-step implementation plan:\n   - Data preparation and input formats\n   - Model selection and loading\n   - Inference workflow to generate frames and ensure temporal coherence\n   - Frame post-processing (stabilization, denoising)\n   - Audio integration (optional) and video encoding (e.g., H.264/HEVC)\n   - Packaging into a reproducible pipeline (config files, environment)\n4) A runnable Python code scaffold (PyTorch or compatible) showing how to load a local model, run inference on a batch of images, assemble frames into a video, and save output to disk. Include a minimal CLI example.\n5) A sample commands/steps to install dependencies, download models, and execute a sample generation run.\n6) Safety, privacy, and ethics: emphasize that all processing is local, respect licenses, avoid harmful content, provide a simple content policy and logging, and outline how to implement offline content moderation through rule-based checks.\n7) Evaluation and QA: suggested metrics (PSNR, SSIM, perceptual quality scores, FPS) and a method to compare results across runs.\n8) Troubleshooting: common issues (model mismatches, memory errors, slow IO) and quick fixes.\n9) References and resources: links to open-source repositories, licenses, documentation.\n10) README skeleton: outline sections and example commands.\n\nDeliver the content in a clear, copy-pasteable format with code blocks where appropriate and concrete figures where feasible. Do NOT provide any guidance that would help bypass safety policies or engage in illicit activity; keep everything self-contained and legally compliant.\n\nIf possible, include a minimal code snippet that demonstrates loading a local model, generating a small sequence of frames, and saving them as an MP4, plus a placeholder for the model path and input directory.