ComfyUI-RH-Bernini

ComfyUI-RH-Bernini is a standalone ComfyUI custom node pack for Bernini-style in-context conditioning on Wan 2.2 workflows. It backports the core runtime behavior from ComfyUI PR #14216 and adds prompt-enhancement helper nodes adapted from the official ByteDance Bernini repository.

This package is not a full standalone port of the official ByteDance Bernini inference pipeline. It is intended for ComfyUI-native Wan 2.2 workflows where model loading, text encoding, sampling, high/low-noise scheduling, and video output are handled by existing ComfyUI nodes.

Author

This custom node package is authored and maintained by flybirdxx.

Code Origin

This code is forked and backported from Comfy-Org/ComfyUI#14216, which adds Bernini in-context conditioning support to ComfyUI. Runtime patch code in this package is derived from that PR so it can be used as a standalone custom node pack before the PR is available in the target ComfyUI runtime.

Nodes

• Bernini Conditioning: encodes source video, reference video, and reference images into context_latents on positive and negative conditioning. The Wan model patch appends those clean latents as in-context tokens with per-stream RoPE source IDs.
• Bernini Prompt Enhancer: builds task-aware prompt-enhancement instructions from the official Bernini prompt enhancer. It outputs separate system_prompt and user_prompt strings for external LLM nodes, plus llm_prompt, api_prompt, and json_mode.
• Bernini Prompt Result Parser: parses the external LLM response into the final enhanced prompt. JSON mode looks for rewritten_text.

Requirements

• Recent ComfyUI with V3 custom node support through comfy_api.latest.
• Wan 2.2 model workflow support in ComfyUI.
• Bernini-compatible Wan 2.2 weights.
• No extra Python packages are required by this node pack beyond the ComfyUI runtime.

Do not install or replace torch, torchvision, torchaudio, CUDA, or NVIDIA packages from this plugin. Use the versions already provided by your ComfyUI environment.

Model Layout

The plugin does not download or load models directly. Use your existing ComfyUI Wan 2.2 model loaders.

Recommended model sources:

• Wan2.2 base model: https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B-Diffusers
• Official Bernini weights: https://huggingface.co/ByteDance/Bernini
• Kijai ComfyUI-compatible Bernini weights: https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/tree/main/Bernini

If you use Hugging Face downloads in a ComfyUI environment, prepare models under your ComfyUI models directory according to the loader nodes used by your workflow.

Example manual preparation commands:

cd /workspace/ComfyUI/models
hf download ByteDance/Bernini --local-dir Bernini

The exact destination path depends on the Wan loader nodes in your workflow.

Usage

1. Build a Wan 2.2 ComfyUI workflow with the Bernini-compatible checkpoint, VAE, text encoder, sampler, and video output nodes.
2. Add Bernini Conditioning under conditioning/video_models.
3. Connect positive, negative, vae, and set width, height, length, and batch_size.
4. Connect optional media inputs according to the task:

| Inputs connected | Task | | --- | --- | | none | t2v | | source_video | v2v | | source_video + reference_images | rv2v | | reference_images only | r2v | | source_video + reference_video | video insertion / ads2v |

5. Use the latent output as the empty latent for sampling.

Prompt Enhancer Workflow

The official Bernini inference scripts recommend prompt enhancement. This plugin does not call any API internally.

Recommended graph:

1. Add Bernini Prompt Enhancer.
2. Connect system_prompt to an external LLM node's system input.
3. Connect user_prompt to the same LLM node's user or message input.
4. For vision tasks, also connect the same source_video, source_image, or reference_images to that LLM node if it supports visual input.
5. Connect the LLM text response to Bernini Prompt Result Parser.
6. Connect json_mode from the enhancer to json_mode on the parser for r2v, r2i, rv2v, and vrc2v.
7. Use the parser's enhanced_prompt output for text encoding.

If your LLM node only supports one prompt input, use llm_prompt instead. api_prompt is kept as the raw official prompt request for compatibility and debugging.

Examples

API-format examples are stored in examples/.

• examples/bernini_prompt_enhancer_api.json: minimal prompt enhancer and parser graph. Replace the placeholder prompt and the placeholder LLM response with your own values or connect the outputs to an external LLM node in a full workflow.

Limitations

• This plugin patches only Wan-related ComfyUI runtime behavior needed by PR #14216.
• It does not implement the official Bernini single-GPU or multi-GPU inference scripts.
• It does not include official model weights or example media.
• It does not provide a built-in LLM API call. Use an external API node.

Source And Related Links

• Project author and maintainer: https://github.com/flybirdxx
• Official ByteDance Bernini repository: https://github.com/bytedance/Bernini
• Code fork and backport source, ComfyUI Bernini PR: https://github.com/Comfy-Org/ComfyUI/pull/14216
• Kijai Bernini branch used by the PR: https://github.com/kijai/ComfyUI/tree/bernini
• Bernini paper: https://arxiv.org/abs/2605.22344
• Bernini project page: https://bytedance.github.io/Bernini
• Official Bernini model weights: https://huggingface.co/ByteDance/Bernini
• Wan2.2 base model: https://huggingface.co/Wan-AI/Wan2.2-T2V-A14B-Diffusers
• Kijai ComfyUI-compatible weights: https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/tree/main/Bernini
• Official ByteDance Bernini license: https://github.com/bytedance/Bernini/blob/main/LICENSE
• ComfyUI license: https://github.com/comfyanonymous/ComfyUI/blob/master/LICENSE

License

This custom node pack is distributed under GPL-3.0-only because it backports and patches ComfyUI runtime code. The prompt-enhancement templates adapted from ByteDance Bernini are attributed to the upstream Apache-2.0 project. See LICENSE for details.