ComfyUI-Bernini

最新版 Bernini 导演台（多段时间轴）请迁移至 AIMixer/ComfyUI_Bernini_Director。

For the latest Bernini Director (multi-segment timeline), please migrate to AIMixer/ComfyUI_Bernini_Director.

Maintenance notice — this repository is no longer actively developed.

For the latest Bernini Director (multi-segment timeline), please migrate to AIMixer/ComfyUI_Bernini_Director.

We have switched fully to the official ComfyUI Bernini-R stack (native VAE / MODEL / CLIP, context_latents, dual-stage KSamplerAdvanced; see PR #14216). Compared with the legacy wrapper pipeline in this repo, the official path is generally more stable, better optimized, and faster in real runs, and it tracks ComfyUI core improvements automatically.

All future maintenance and feature work will land only on the official-path project. This repository remains available as a historical reference.

维护说明：本仓库已停止更新。

最新版 Bernini 导演台（多段时间轴）请迁移至 AIMixer/ComfyUI_Bernini_Director。

我们已全面改用 ComfyUI 官方 Bernini-R 技术路径 （原生 VAE / MODEL / CLIP、context_latents、双阶段 KSamplerAdvanced； 详见 PR #14216）。 相较本仓库的独立 wrapper 旧链路，官方路径通常更稳定、优化更完善、实际运行也更快， 并能随 ComfyUI 核心一并获得持续改进。

后续功能与修复将全部基于官方技术路径维护，仅在新仓库发布。 本仓库仅保留作历史参考。

Standalone ComfyUI plugin with a complete Wan 2.2 Bernini pipeline that does not depend on ComfyUI core Bernini nodes (PR #14216).

中文文档 → README_ZH.md

Quick start

1. Clone into ComfyUI/custom_nodes/:

   cd ComfyUI/custom_nodes
   git clone https://github.com/AIMixer/ComfyUI-Bernini.git

2. Install Python dependencies (use the same Python environment as your ComfyUI):

   cd ComfyUI-Bernini
   pip install -r requirements.txt

   On Windows portable builds, prefer ComfyUI's bundled Python, for example:

   ..\..\python_embeded\python.exe -m pip install -r requirements.txt

3. Restart ComfyUI. Nodes appear under the Bernini category.

4. Download models — quantized GGUF / FP8 + workflows: comfyit.cn/article/489; original Kijai FP8 only: HuggingFace Bernini. Put weights under ComfyUI/models/ (diffusion_models/, vae/, text_encoders/). See Quantized models.

5. Run a workflow: Model Loader → Context Embeds → Sampler → Decode. Example JSON workflows: Comfyit article 489.

Quantized models

Download: Comfyit article 489 — Bernini models & workflows (GGUF quantizations, scaled FP8, VAE, T5, and example JSON).

Wan 2.2 Bernini diffusion weights are available in GGUF and FP8 safetensors quantizations. Place files under ComfyUI/models/diffusion_models/ and load both HIGH and LOW in Bernini Model Loader (GGUF is supported).

Tier	GGUF (LOW / HIGH)	Min VRAM (Bernini Director)
Q4_K_M (lowest)	Wan22_Bernini_LOW-Q4_K_M.gguf · Wan22_Bernini_HIGH-Q4_K_M.gguf	8 GB
Q5_K_M	Wan22_Bernini_LOW-Q5_K_M.gguf · Wan22_Bernini_HIGH-Q5_K_M.gguf	10 GB
Q6_K	Wan22_Bernini_LOW-Q6_K.gguf · Wan22_Bernini_HIGH-Q6_K.gguf	12 GB
Q8_0	Wan22_Bernini_LOW-Q8_0.gguf · Wan22_Bernini_HIGH-Q8_0.gguf	16 GB

FP8 safetensors (scaled, same naming as Kijai Bernini pack):

• Wan22_Bernini_LOW_fp8_e4m3fn_scaled.safetensors
• Wan22_Bernini_HIGH_fp8_e4m3fn_scaled.safetensors

VRAM tips (Bernini Director, HIGH + LOW): Q4 → 8 GB min; Q5 → 10 GB; Q6 → 12 GB; Q8 → 16 GB. Enable Block Swap on both loaders; use T5 disk cache or fp8 text encoder where possible.

All quantized builds above are included in the resource pack at comfyit.cn/article/489. For the original non-GGUF FP8 pack from Kijai, see HuggingFace.

Node chain

BerniniModelLoader · BerniniVAELoader · BerniniTextEncodeCached · BerniniContextEmbeds · BerniniContextOptions · BerniniSamplerExtraArgs · BerniniScheduler · BerniniSampler · BerniniDecode · BerniniDirector · BerniniDirectorOfficial

Bernini Director (KJ backend)

All-in-one node with an embedded timeline editor: upload video and reference images inside the node, split segments, set per-segment prompts / task_type, then run the full Bernini HIGH/LOW pipeline in one queue. Uses Bernini Model Loader (GGUF + Block Swap) for lower VRAM.

Example workflows: see Example workflows below (all from Comfyit article 489).

Bernini Director Official (ComfyUI core backend)

Same timeline UI as Bernini Director, but execution goes through ComfyUI native Bernini (VAELoader / UNETLoader ×2 / CLIPLoader → BerniniConditioning → dual-stage KSamplerAdvanced), aligned with PR #14216.

	KJ Director	Official Director
Models	Bernini Model Loader (GGUF)	UNETLoader fp8_scaled
Text	Bernini Text Encode Cached	CLIPLoader (wan)
VRAM	~8 GB with Q4 GGUF	Higher; use --lowvram, CLIP on CPU

Typical v2v workflow: load bernini_director_official_core_v2v_2.json → connect VAE / UNET×2 / CLIP → upload source video → Equal Split into segments → global prompt → queue once, export all segments.

Example workflows

Download Bernini model weights + example JSON workflows from Comfyit: Bernini models & workflows (article 489):

Workflow	task_type	Download
bernini_director_minimal_test (r2v) .json	r2v	comfyit.cn/article/489
bernini_director_minimal_test (t2i) .json	t2i	comfyit.cn/article/489
bernini_director_minimal_test (t2v) .json	t2v	comfyit.cn/article/489
bernini_director_minimal_test (r2i) .json	r2i	comfyit.cn/article/489
bernini_director_minimal_test (v2v).json	v2v	comfyit.cn/article/489
bernini_director_minimal_test (i2v) .json	i2v	comfyit.cn/article/489
bernini_director_minimal_test (i2i).json	i2i	comfyit.cn/article/489
bernini_director_minimal_test (rv2v).json	rv2v	comfyit.cn/article/489
bernini_director_minimal_test (vi2v) .json	vi2v	comfyit.cn/article/489
bernini_director_minimal_test (vrc2v) .json	vrc2v	comfyit.cn/article/489
bernini_director_minimal_test (mv2v) .json	mv2v	comfyit.cn/article/489
bernini_director_minimal_test (ads2v).json	ads2v	comfyit.cn/article/489
bernini_director_official_core_v2v_1.json	v2v	comfyit.cn/article/489
bernini_director_official_core_v2v_2.json	v2v	comfyit.cn/article/489
bernini_director_official_core_rv2v.json	rv2v	comfyit.cn/article/489
bernini_video_edit(r2v) .json	r2v	comfyit.cn/article/489
bernini_video_edit(v2v).json	v2v	comfyit.cn/article/489
bernini_video_edit(vi2v) .json	vi2v	comfyit.cn/article/489
bernini_video_edit(rv2v) .json	rv2v	comfyit.cn/article/489

After download: merge models/ into ComfyUI/models, install plugins/deps, drag the JSON into ComfyUI. Details on the article page.

Acknowledgements

The engine/ layer is adapted from kijai/ComfyUI-WanVideoWrapper (Apache-2.0) and wuwukaka/ComfyUI-WanAnimatePlus. Deep respect and gratitude to kijai and all contributors to the WanVideo ecosystem.

License

This project is licensed under the Apache License, Version 2.0.

The engine/ layer is adapted from kijai/ComfyUI-WanVideoWrapper (also Apache-2.0) and wuwukaka/ComfyUI-WanAnimatePlus.. See Acknowledgements.

---

Ecosystem · Comfyit 搅拌站

Comfyit is a one-stop ComfyUI tools & learning platform. For environment setup, models, workflows, and tutorials that complement this plugin, see the Product Center (ComfyUI Manager, LoRA Trainer, Prompt Master) and free resources: packages · models · workflows · learning center.

Full details in README_ZH.md.

Contact

Maintainer	AIMixer
Author QQ	3697688140
Bilibili	space.bilibili.com/1997403556
QQ groups	551482703 · 425064221 · 559826331
Comfyit	comfyit.cn