Bernini is a latent semantic planning framework for video and image generation and editing, developed by ByteDance. It unifies tasks such as text-to-image (t2i), image-to-image (i2i), text-to-video (t2v), video-to-video (v2v), reference-driven video generation (r2v/rv2v), and ad insertion (ads2v) under a two-tier architecture. The full Bernini pipeline uses a Qwen2.5-VL multi-modal large language model as a 'semantic planner' to produce latent feature plans, which are then decoded by a connector and a MaskGIT-style iterative VIT decoder into conditions for a dual-expert Wan2.2 diffusion renderer. A lighter variant, Bernini-R, omits the planner and fine-tunes the renderer directly with a T5 encoder. Inference is supported via single-GPU and multi-GPU scripts (with Ulysses sequence parallelism), a Gradio demo, and JSON-driven case files that define task configuration, prompts, and source media. Training of the renderer uses the VeOmni distributed framework with FSDP2 and Ulysses SP, and leverages pre-extracted VAE latents and VIT embeddings stored in Parquet format. The codebase is published under Apache 2.0 with pretrained models on Hugging Face, accompanied by a paper and benchmark results on EditVerse, OpenVE, OpenS2V, VBench, and a custom arena.