WAN 2.2 Bernini Video Editing Image-to-Video Model ComfyUI

The newly released WAN 2.2 Bernini Video Edit Model represents a major step forward in open-source AI video generation, specifically focused on high-quality image-to-video and video-to-video editing workflows. Built on a 14B parameter diffusion architecture, Bernini is designed to produce temporally consistent, cinematic-quality outputs with strong motion coherence and improved prompt adherence compared to earlier WAN releases. It supports both high-noise and low-noise pipelines, allowing for flexible control over motion intensity and detail preservation during video edits.
 
In terms of capabilities, WAN 2.2 Bernini excels at transforming still images into dynamic video sequences, refining existing footage, and applying stylistic transformations while maintaining structure and subject identity. The model supports LoRA-based acceleration workflows such as Lightx2v, enabling faster generation with fewer steps while still achieving high-quality outputs.
Running WAN 2.2 Bernini locally does require significant compute, especially for full FP16 models, which can demand 48GB+ VRAM for optimal performance.
 
However, this setup has been specifically optimized to run on 24GB VRAM GPUs (such as RTX 3090, RTX 4090, or RTX A5000) through the use of GGUF quantized UNET models, FP8 text encoders, and efficient LoRA-based workflows. With further tweaks—such as lowering resolution, reducing frame count, or adjusting sampling steps—it is possible to run on slightly lower VRAM configurations, though with trade-offs in quality and speed.
 
To make this accessible, I’ve included a one-click installer alongside a customized version of Kijai’s ComfyUI workflow. This setup removes much of the manual configuration typically required, preloading key dependencies and structuring the workflow for efficient memory usage. The result is a streamlined experience that allows both beginners and advanced users to quickly start generating AI video edits using WAN 2.2 Bernini without needing to manually troubleshoot node setups or compatibility issues.
 
Preloaded Models Within the Installer (Low VRAM)
umt5_xxl_fp8_e4m3fn_scaled.safetensors (ComfyUI\models\clip)
Hugging Face Link
wan_2.1_vae.safetensors (ComfyUI\models\vae)
Hugging Face Link
bernini_r_high_noise_14B-Q4_K_M.gguf (ComfyUI\models\unet)
Hugging Face Link
bernini_r_low_noise_14B-Q4_K_M.gguf (ComfyUI\models\unet)
Hugging Face Link
high_noise_model.safetensors Lightx2v LoRA (ComfyUI\models\loras)
Hugging Face Link
low_noise_model.safetensors Lightx2v LoRA (ComfyUI\models\loras)
Hugging Face Link
 
The standard WAN 2.2 Bernini 14B diffusion models (FP16 and FP8) are not bundled with this installer, but you can easily grab them from the official Comfy Org Hugging Face repository and drop them into your ComfyUI/models/diffusion_models folder.
Comfy Org WAN 2.2 Bernini Diffusion Models
 
Custom Node w/ command needed for Portable ComfyUI Windows package(Manual Install)
ComfyUI Manager - https://github.com/ltdrdata/ComfyUI-Manager
Command to install requirements:
.\python_embeded\python.exe -m pip install -r .\ComfyUI\custom_nodes\ComfyUI-Manager\requirements.txt
 
Use the ComfyUI Manager’s Install missing custom nodes feature to install any additional nodes that may not work initially after installation. In testing, a few nodes only installed correctly through the manager.
 
Speed
Edit 10-second 480p videos (6 steps) in under 10 minutes including upscaling on an RTX A5000 (24GB VRAM).
The workflow scales with your hardware, allowing 720p+ and higher frame rates (24fps+) on stronger GPUs.
 
System Requirements
Nvidia RTX 30XX / 40XX / 50XX GPU (FP16 supported)
CUDA-compatible GPU (minimum 24GB VRAM)
Windows OS
Minimum 40GB free storage
 
What’s Included
Portable ComfyUI Windows Installer fully configured for WAN 2.2 Bernini image-to-video workflows
Automated downloads for required models and custom nodes
Optimized workflow for both beginners and advanced users
Fast Groups Bypasser support for quickly enabling/disabling workflow sections
 
Usage Notes
Download and place the installer files in a dedicated folder, then double-click the installer to set up ComfyUI and the WAN 2.2 Bernini workflow automatically (no manual configuration required).
Load the included ComfyUI workflow JSON file after installation to access the preconfigured WAN 2.2 Bernini image-to-video pipeline.
Upload your reference video (source footage) along with the reference image you want to integrate or guide the edit. This workflow is optimized for image-to-video and video-to-video transformations.
You can use either the WAN 2.2 Bernini 14B GGUF models (recommended for 24GB VRAM and lower memory usage) or the standard WAN 2.2 Bernini 14B diffusion models (FP16/FP8) for higher quality if you have more VRAM available.
Use the Fast Groups Bypasser node (left sidebar) to quickly enable or disable sections of the workflow, including LoRA usage, upscaling, or alternative render paths.
For best results, write detailed and consistent prompts for each 5-second segment. Maintaining subject, style, and motion continuity between segments will significantly improve temporal consistency and reduce flickering.
Start with lower resolutions such as 480p when testing settings, then scale up to 720p or higher once you find stable configurations for your GPU.
Adjust sampling steps, frame count, and denoise strength to balance quality, speed, and VRAM usage. Lower steps (4–6) with Lightx2v LoRAs provide fast results, while higher steps improve detail at the cost of render time.
If you encounter VRAM limitations, reduce resolution, batch size, or disable optional nodes like upscaling and secondary passes.
Keep your output clips short (5–10 seconds per segment) and stitch them together in post for longer videos. This improves stability and reduces generation errors.
Always verify model paths inside ComfyUI if something fails to load, especially when adding external WAN 2.2 Bernini diffusion models manually.