Bernini v2 Comes to ComfyUI via Custom Nodes

Overview

The full Bernini v2 pipeline was previously only runnable through ByteDance's official diffusers code. The ComfyUI pack converts it into a node-based workflow: a fine-tuned Qwen2.5-VL planner with a connector and a MaskGIT-style ViT decoder turns text plus source visuals into semantic plan tokens, then two co-trained Wan2.2 DiTs (high-noise and low-noise, switched at 0.875) render the video with source media injected as context latents.

The Bernini v2 planning and rendering pipeline running as ComfyUI custom nodes. Source: rzgar/Bernini-v2-ComfyUI

Six Tasks, One Workflow

Task selection is automatic based on which media inputs are connected:

Recommended settings: CFG 3, 16 to 50 steps, UniPC / DPMPP_2M / EULER samplers.

Model Files

The pack ships FP8-scaled and NVFP4 quantizations of both DiTs, plus the planner assets (MLLM, connector, ViT decoder, mask tokens) that go under models/text_encoders/. The planner tokenizer and processor configs are provided as a separate zip in the repo. All components are listed on the model page.

Background: Bernini v2

ByteDance open-sourced the full Bernini framework in June 2026 and packaged it as Bernini-Diffusers-v2 on August 13. The v2 training recipe warms up the connector for thousands of steps before co-training, which improved reference-guided editing and OpenS2V scores (63.83) over the first diffusers release. In ByteDance's internal blind arena, Bernini ranks in the first tier of video editing among closed-source commercial models.

The Bernini framework: the MLLM semantic planner decomposes instructions, the Wan2.2 DiT renderer produces the video. Source: Bernini project page