Bernini is ByteDance's open-source unified framework for video generation and editing, built around a distinctive two-stage architecture: an MLLM-based semantic planner that decomposes complex instructions and plans semantic changes in latent space, paired with a DiT-based renderer that turns those plans into pixels. Introduced in the paper "Latent Semantic Planning for Video Diffusion" (May 2026) and fully open-sourced across June and July, the repository has passed 1,100 GitHub stars under an Apache-2.0 license, with weights on Hugging Face and — as of July 13 — the full training code for the renderer. ## Planner + Renderer, Not Just a Bigger Diffusion Model Most open video models are monolithic diffusion transformers that must parse an instruction and render it in one shot. Bernini splits the job: a 7B multimodal LLM planner first reasons about what should change semantically — objects, styles, motions, scene logic — and hands a latent plan to a 14B DiT renderer fine-tuned from the Wan lineage. The result is measurably stronger instruction following on compound edits ("remove the watermark, change the season to winter, and keep the camera motion") than renderer-only baselines. On ByteDance's self-built human-preference arena, aggregated with Bradley-Terry scoring over blind pairwise votes, Bernini reaches the first tier among leading closed-source commercial video editors. ## Two Model Families, One Task Interface The repository ships two families that share the same task interface. The full Bernini pipeline (7B planner + 14B renderer) maximizes instruction-following quality, scoring 8.02 on EditVerse and 4.03 on OpenVE. Bernini-R is the renderer alone — simpler to run, with 14B and 1.3B checkpoints. The 1.3B variant, fine-tuned from Wan2.1-1.3B, performs close to the 14B model on simple tasks like style transfer, subtitle/watermark removal, and local editing, making it a practical option for smaller GPUs, while lagging on harder tasks such as human generation. All models handle six task modes: text-to-image, image-to-image, text-to-video, video-to-video, reference-guided video editing (rv2v), and reference-to-video (r2v). ## Open Weights and Training Code ByteDance has been unusually complete about the release cadence: the paper landed May 22, Bernini-R inference code and 14B weights on June 1, the 1.3B checkpoint on June 9, the full planner+renderer pipeline on June 11, and the Bernini-R training code with a full training guide on July 13. Everything is packaged in Diffusers-compatible format on the ByteDance Hugging Face org, which lowers the integration barrier for teams already building on the diffusers ecosystem. ## Benchmark Standing Across public benchmarks the released models post competitive numbers: 84.6+ on VBench for both renderer sizes, 62.18–62.94 on OpenS2V, and steady gains from 1.3B → 14B → full pipeline on the project's own open-sourced v2v evaluation sets. The more interesting signal is the arena result against commercial systems — open-weight video editing that human raters place alongside closed products has been rare. ## Trade-offs and Limitations The full pipeline is heavy: a 7B planner plus a 14B renderer puts serious VRAM demands on inference, and video diffusion remains slow relative to image work. The 1.3B renderer trades capability for accessibility and visibly struggles with human generation. The planner stage adds latency and another failure mode — a wrong semantic plan renders faithfully wrong. Training code currently covers only the renderer (Bernini-R), not the planner, and the arena leaderboard is self-built by ByteDance, so independent replication of the human-preference claims is still pending. ## Who Should Use This Bernini is aimed at researchers studying instruction-following in video diffusion, and at developers building video editing tools who want open weights that approach commercial editing quality. The 1.3B renderer is the entry point for experimentation on a single consumer GPU; the full pipeline is for teams with serious hardware chasing state-of-the-art open video editing. With Apache-2.0 licensing, Diffusers packaging, and now-public training code, it is one of the most complete open video-editing stacks released this year.