rzgar/Bernini-v2-ComfyUI

Bernini v2 is a video editor: t2v / i2v / v2v / r2v / rv2v / ads2v. The source media (reference images / source video) are first-class inputs. The pipeline has two stages:

1. 1.Semantic planning: a fine-tuned Qwen2.5-VL (mllm) + connector + vit_decoder (MaskGIT-style loop) turn the text + source visual tokens into diff_mllm_contexts (the novel part Bernini-R lacks).
2. 2.Rendering: two co-trained Wan2.2 DiTs (high-noise / low-noise, switched at 0.875) with the Bernini vae_txt_vit_wapg guidance (4-way CFG with APG projection), VAE-encoded source media as context_latents, Wan2.1 VAE decode.

ComfyUI Custom node

Extract the ZIP file in ComfyUI/custom_nodes <img src="https://huggingface.co/rzgar/Bernini-v2-ComfyUI/resolve/main/workflow/customnodescreenshot.png" width="1280" height="720" />

Task is inferred from which media inputs are connected: (none) = t2v, source_video = v2v, source_video + reference_images = rv2v, reference_images only = r2v, source_video + reference_video = ads2v.

Recommended settings

• —CFG: 3
• —Steps: 16, 23 or 50
• —Sampler: UniPC / DPMPP_2M / EULER

Model files

• —FP8_scaled

Place the planner files anywhere under `models/text_encoders/` - folder names don't matters: