Trying out the Bernini-R-LightX2V-4step-loras for Wan2.2 Bernini in ComfyUI

A LoRA for Wan2.2 Bernini has been released, so I will give it a try.

The LoRA I used when introducing it in the article below was for T2V, so it seems this one was created specifically for Bernini.

Also, since the memory requirements are tight and difficult to work with, I have prepared a BlockSwap node to offload processing to CPU memory.
If you don't use this, processing that overflows from VRAM will spill into shared GPU memory, making the process hopelessly slow.

Originally, this was available in Kijai's custom nodes for Wan, but it only supported T2V and I2V and was not compatible with the combination of Bernini and official ComfyUI nodes, so I created one.

For reference, I will also list my previous articles related to Wan2.2.

TI2V

I2V

I2V (VBVR LoRA)

FLF

ControlNet

Long-form

TTM

V+V2V (Video concatenation)

Video editing (Bernini)

I also explain it in a video.

Environment

OS: Windows 11
GPU: GeForce RTX 4090
CPU: i9-13900KF
memory: 128G

Setup procedure

Please install the following custom nodes and update ComfyUI to the latest version.

cd ComfyUI\custom_nodes
git clone https://github.com/lovemachine100/ComfyUI-JITBlockSwap.git

Also, please add the "--disable-dynamic-vram" and "--disable-pinned-memory" options to your ComfyUI startup bat file.

ComfyUI_windows_portable\run_nvidia_gpu.bat

Before change

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build

After change

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-dynamic-vram --disable-pinned-memory

--disable-dynamic-vram: A flag that disables dynamic VRAM management by comfy-aimdo (enabled by default) and reverts to the traditional legacy ModelPatcher method.

--disable-pinned-memory: A flag that completely prohibits the use of pinned memory (page-locked host memory).

*Please remove these when not using BlockSwap. *It is recommended to create a separate run_nvidia_gpu.bat from your existing one.

Models

wan2.2_bernini_r_high_noise_fp16.safetensors and wan2.2_bernini_r_low_noise_fp16.safetensors downloaded and placed in "ComfyUI\models\diffusion_models"

Bernini-R_LightX2V_high_noise.safetensors and Bernini-R_LightX2V_low_noise.safetensors downloaded and placed in "ComfyUI\models\loras"

wan_2.1_vae.safetensors downloaded and placed in "ComfyUI\models\vae"

umt5_xxl_fp8_e4m3fn_scaled.safetensors downloaded and placed in "ComfyUI\models\text_encoders"

Workflow

The workflow is as follows.

Regarding Block Swap, increasing the number for blocks_to_swap will decrease the percentage of VRAM used and increase the percentage of CPU memory used.

With the current settings, a value of around 20 will use about 24GB of VRAM.
At around 30, VRAM usage is kept to about 18GB. (CPU memory is used instead, and processing speed will be slower.)

However, if VRAM usage exceeds 24GB, processing will start to spill over into shared GPU memory, causing data transfer overhead and making processing performance desperately slow.

Therefore, it is optimal to aim for a blocks_to_swap value that keeps VRAM usage just under 24GB. (In this case, it would be around 25, I suppose.)

Generation Results

The generation results are as follows.
Generation time: 420 seconds.
CPU memory: 100GB
GPU memory: 18GB VRAM

Compared to the results from the previous article, the twin tails are now reflected, so the quality has improved. Thanks to the Block Swap effect, I was able to increase the frame count while also improving the generation speed.

Impressions

The fact that both quality and generation speed were improved is a big plus this time.

Personally, I believe that the demand for video generation AI for video editing purposes will increase as the cost of video generation rises, so I hope to gradually increase my options.