ComfyUI / RunningHub wrapper for the full Bernini pipeline. The plugin uses ByteDance/Bernini-Diffusers only and loads the local Bernini-Diffusers-qint8 model directory by default. It does not load Bernini-R, high_noise_ckpt, or low_noise_ckpt.

The nodes use standard ComfyUI IMAGE and VIDEO sockets, so they can be connected directly to LoadImage, LoadVideo, SaveImage, and SaveVideo. The old all-parameter node has been removed; internal model paths and output prefixes are no longer exposed as node widgets.

Features

Text to image: RHBerniniFullTextToImage
Image editing: RHBerniniFullImageToImage
Text to video: RHBerniniFullTextToVideo
Video editing: RHBerniniFullVideoToVideo
Reference image to video: RHBerniniFullReferenceToVideo
Reference image + video editing: RHBerniniFullReferenceVideoToVideo

Installation

Install inside your ComfyUI Python environment. Do not use this plugin's dependencies to replace the torch / CUDA packages already provided by your ComfyUI runtime.

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-Bernini-Full.git
cd ComfyUI-RH-Bernini-Full
pip install -r requirements.txt
pip install --no-deps git+https://github.com/ByteDance-Seed/[email protected]

Model Download & Installation

The recommended runtime model is the pre-quantized qint8 package:

Model Directory Structure

Place the model under ComfyUI's diffusers model directory:

ComfyUI/
└── models/
    └── diffusers/
        └── Bernini-Diffusers-qint8/
            ├── bernini/
            ├── mllm/
            ├── t5_text_encoder/
            ├── t5_tokenizer/
            └── vae/

Download Methods

Method 1: Download from ModelScope

cd ComfyUI/models
modelscope download --model Gluttony10/Bernini-Diffusers-qint8 --local_dir diffusers/Bernini-Diffusers-qint8

The plugin looks for this directory by default:

ComfyUI/models/diffusers/Bernini-Diffusers-qint8

The node validates that config.json has model_type: "bernini" and rejects paths containing Bernini-R or Bernini_R. The unquantized bytedance-community/Bernini-Diffusers package is useful for local conversion or debugging; to run it directly, keep the full Bernini-Diffusers structure and place it at the default path or create a same-name symlink.

Method 2: Manual Download

| Model | Link | Description | |-------|------|-------------| | Bernini-Diffusers-qint8 | https://modelscope.cn/models/Gluttony10/Bernini-Diffusers-qint8 | Recommended qint8 runtime package | | Bernini-Diffusers | https://huggingface.co/ByteDance/Bernini-Diffusers | Upstream full model for conversion or debugging | | Bernini-Diffusers | https://modelscope.cn/models/bytedance-community/Bernini-Diffusers | ModelScope mirror of the upstream full model |

Common Parameters

quality_preset: only standard and quality are available. standard is the recommended default; quality spends more planning time.
acceleration: defaults to wan2.2_lightx2v_4step. If no matching Wan2.2 T2V LightX2V high/low-noise LoRA pair is found under ComfyUI/models/loras, the node falls back to inference without LoRA.
steps: official reproduction workflows use the original script step counts; with LightX2V acceleration enabled, 4 steps is the expected fast setting.
memory_mode: defaults to keep_loaded_batch, allowing consecutive t2i/i2i/v2v/r2v jobs to reuse the same cached Bernini pipeline. Use low_vram_unload on memory-constrained machines.
width / height: image and text-to-video nodes control output size directly; video editing nodes follow the source video when both are set to 0.

Memory And Speed

The model cache is kept by default, so repeated runs with the same model and LoRA configuration avoid another cold load.
low_vram_unload releases VAE, MLLM/VIT, T5, and diffusion transformer stages in order after each run, then clears CUDA cache.
VAE tiling / slicing is enabled internally. Multi-reference images are encoded in chunks to reduce peak VRAM.
More VRAM mainly allows larger resolution, more frames, and less unloading. It does not remove the first qint8 model load cost.

Wan2.2 LightX2V LoRA

The node only searches the currently adapted high/low-noise LoRA pair under ComfyUI/models/loras. Other similarly named LightX2V / I2V LoRAs are not auto-loaded because they are not a good fit for the full Bernini pipeline and can noticeably reduce quality. All task types use this T2V LoRA pair so they can share the cached Bernini pipeline.

ComfyUI/models/loras/Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V2.0/high_noise_model.safetensors
ComfyUI/models/loras/Wan2.2-T2V-A14B-4steps-lora-rank64-Seko-V2.0/low_noise_model.safetensors

Example Workflows

ComfyUI API prompt examples are available in examples/. Their class_type values use the plugin registration names and can be submitted to /prompt or test runners. Assets and parameters follow bytedance/Bernini assets/testcases/* and scripts/bernini/run_*.sh.

bernini_t2i_test_api.json: text to image, upstream t2i/t2i.json.
bernini_i2i_test_api.json: image editing, upstream i2i/i2i.json.
bernini_t2v_test_api.json: text to video, upstream t2v/t2v.json.
bernini_v2v_test_api.json: video editing, upstream v2v/v2v_case1.json.
bernini_r2v_test_api.json: multi-reference image to video, upstream r2v/r2v.json.
bernini_rv2v_test_api.json: reference image + video editing, upstream rv2v/rv2v_case1.json.

Before running media-input examples, copy the bundled assets to your ComfyUI input directory:

cp custom_nodes/ComfyUI-RH-Bernini-Full/examples/assets/bernini_i2i_source.png input/bernini_i2i_source.png
cp custom_nodes/ComfyUI-RH-Bernini-Full/examples/assets/bernini_v2v_source_case1.mp4 input/bernini_v2v_source_case1.mp4
cp custom_nodes/ComfyUI-RH-Bernini-Full/examples/assets/bernini_rv2v_source_case1.mp4 input/bernini_rv2v_source_case1.mp4
cp custom_nodes/ComfyUI-RH-Bernini-Full/examples/assets/bernini_rv2v_ref_case1.jpg input/bernini_rv2v_ref_case1.jpg
cp custom_nodes/ComfyUI-RH-Bernini-Full/examples/assets/bernini_r2v_source_img0.png input/bernini_r2v_source_img0.png
cp custom_nodes/ComfyUI-RH-Bernini-Full/examples/assets/bernini_r2v_source_img1.png input/bernini_r2v_source_img1.png
cp custom_nodes/ComfyUI-RH-Bernini-Full/examples/assets/bernini_r2v_source_img2.png input/bernini_r2v_source_img2.png
cp custom_nodes/ComfyUI-RH-Bernini-Full/examples/assets/bernini_r2v_source_img3.png input/bernini_r2v_source_img3.png
cp custom_nodes/ComfyUI-RH-Bernini-Full/examples/assets/bernini_r2v_source_img4.png input/bernini_r2v_source_img4.png

See examples/README_CN.md for more details.

Limitations

Only full Bernini-Diffusers is supported; Bernini-R is intentionally unsupported.
The plugin performs local inference and does not download large models or LoRAs automatically.
Video jobs are slow and VRAM-heavy. For environment checks, start with the text-to-image node.

License

This project is released under the Apache License 2.0. See LICENSE for details.

Acknowledgements

This project is based on Bernini and Bernini-Diffusers from ByteDance.

ComfyUI-RH-Bernini-Full

Nodes (6)