Video - https://youtu.be/3-fVYx3Vc-s

Overview

Bernini-R is now available using ComfyUI and the native Bernini Conditioning node! It uses WAN 2.2 to provide both image and video editing tasks along with a range of other features such as text or image to video.

Huggingface: https://huggingface.co/ByteDance/Bernini-R

Project Page: https://bernini-ai.github.io/

Paper: https://arxiv.org/abs/2605.22344

GitHub Code: https://github.com/bytedance/Bernini

Various models are available, but the workflow I've been building and testing uses the ComfyUI model conversions available at https://huggingface.co/Comfy-Org/Bernini-R

They have both fp8 and fp16 model versions available, and ideally you'll have 16GB+ VRAM for the best performance.

It also looks to be possible with just 4GB VRAM as a minimum using the 1.3B model - at least according to https://github.com/neuregex/ComfyUI-BerniniR, though I've not used their nodes, but for super low VRAM users it looks like one optional alternative. The original Bernini-R GitHub repo also provides an alternative to using ComfyUI.

Workflow Information

This workflow is jam-packed with features including a 2x image upscaler and video interpolation for higher fps results.

Custom Nodes

The workflow mostly uses native ComfyUI Nodes, but some custom nodes are required for extra features. The note in the workflow provides links to the custom nodes required:

Model Downloads

If you've used WAN2.2 before, you'll have most of the files already - though the Bernini-R model files are new:

Personally I'm using the Seko V2 LoRAs (from the See Also: links section), but the lightx2v_T2V-14B used by Comfy is also good. Most of the lightx2v LoRAs seem to work just fine, I simply think the Seko V2 ones have been giving me slightly better image and video quality. Your mileage may vary!

Model Loader

The Model Loader has boat-loads of options, and the defaults are what I think give pretty reasonable results:

Being WAN 2.2 it is super slow without using the lightx2v LoRAs, so I've not really used it much without because it simply takes too long.

If you don't want to use the LoRAs, you can flip the switches, but remember you'll also need to change the CFG to 5 and steps to at least 40:

You can also disable SkipLayerGuidance if you like.

To help reduce VRAM requirements, you can also enable sage attention (if you have it installed) and use Chunk FeedForward:

The Sage Attention GitHub page is at https://github.com/thu-ml/sageattention

Version 2.2.0 worked for me using their instructions:

git clone https://github.com/thu-ml/SageAttention.git

cd SageAttention

export EXT_PARALLEL=4 NVCC_APPEND_FLAGS="--threads 8" MAX_JOBS=32 # Optional

python setup.py install

Microsoft Windows users may prefer to use something like https://github.com/woct0rdho/SageAttention - or just not enable the option ;)

Prompting

For prompting you can use a wide variety of styles, the default being for Image Editing:

As the model can also do text-2-image, text-2-video, image-2-video and a whole lot more, then your style of prompt will change depending on what you want to do and the inputs you select.

Options

There are a lot of options! In the custom combo node, you can choose from the 12 task types. Default actually works fine for most things, as all this node does is select some prefix text for your prompt:

The text it uses is up in the "dragons" section, for those who like the technical stuff:

Also, I've found that using the "default" mode actually work just fine in the vast majority of cases. There probably are cases where changing the choice in the combo box really helps, but I have yet to actually find them...

The length should be set to 1 for image editing tasks, and is the number of frame for video editing ones.

Other options include steps, width, height, seed and fps:

Note that the width and height are only used for tasks which don't use a source / reference video or input image0.

The workflow default is for Image Editing, and so the groups enabled are image0, save image (and 2x) and image compare:

You should change these depending on the task you're looking to do!

Image Editing

When doing Image Editing, the size of Image0 is used for the final result:

Unlike Qwen Image Edit, the input doesn't shift around:

You can change styles, add / remove objects, etc - all with decent results:

Video Editing

Video Editing work in much the same way, and the resolution you set in the Load Video node is the one used for generations:

The default FPS is 126, but for video tasks you'll want to change that in the Options:

Use a simple prompt such as "Change the man's hair colour to black", you can get an edited video!

Edit Video with Reference (Face Swap)

You can also use reference images for video editing tasks:

This will, of course, produce perfectly normal results...

Yes, we're limited to just a few seconds - I've done up to around 10s so far - but then you can always use my LTX video workflow to extend or add sound!

Have fun!