原生 ComfyUI 节点，不依赖 Diffusers 运行时。
六种官方任务和可直接打开的示例工作流。
自动低显存调度、顺序 CFG/APG 和双专家切换。
Core 风格的单体 safetensors：Planner、UMT5、高噪声 Wan、低噪声 Wan 各一个文件，不需要 Diffusers 目录、分片索引或 manifest。
BF16 与原生 ComfyUI INT8 ConvRot 权重。
可识别 NVFP4、现代/旧版 scaled FP8；GGUF 通过外部 ComfyUI-GGUF 接入两个 Wan renderer，Planner 仍使用原生权重。
提供低内存分片转 GGUF、原子量化/5D 回填和双专家合同校验工具。

安装节点

在 ComfyUI-Manager 中搜索 Bernini v2 (Native)，或手动安装：

Registry 版本 0.3.3 已通过审核，当前状态为 Active。

cd C:/path/to/ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-bernini-v2-T8.git

重启 ComfyUI。

下载模型

模型仓库：https://huggingface.co/t8star/Bernini-V2-Comfy

推荐的 Balanced INT8：

hf download t8star/Bernini-V2-Comfy `
  --include "text_encoders/*_int8.safetensors" "diffusion_models/*_int8.safetensors" "vae/wan_2.1_vae.safetensors" `
  --local-dir C:/path/to/ComfyUI/models

BF16 参考版本：

hf download t8star/Bernini-V2-Comfy `
  --include "text_encoders/*_bf16.safetensors" "diffusion_models/*_bf16.safetensors" "vae/wan_2.1_vae.safetensors" `
  --local-dir C:/path/to/ComfyUI/models

实验性 Q4_K_S GGUF（先安装 ComfyUI-GGUF）：

hf download t8star/Bernini-V2-Comfy `
  --include "Bernini-v2-GGUF-Q4_K_S/*" `
  --local-dir C:/path/to/ComfyUI/models/diffusion_models

同一模型仓库还提供经过哈希核对的标准 Wan 2.1 VAE。直接下载到 ComfyUI 的模型根目录：

hf download t8star/Bernini-V2-Comfy `
  --include "vae/wan_2.1_vae.safetensors" `
  --local-dir C:/path/to/ComfyUI/models

也可以使用仓库内固定版本的下载脚本：

python tools/download_vae.py --output C:/path/to/ComfyUI/models/vae

最终目录应类似：

ComfyUI/models/
├── text_encoders/
│   ├── bernini_v2_planner_int8.safetensors
│   └── umt5_xxl_bernini_v2_int8.safetensors
├── diffusion_models/
│   ├── bernini_v2_high_noise_int8.safetensors
│   └── bernini_v2_low_noise_int8.safetensors
└── vae/
    └── wan_2.1_vae.safetensors

打开 examples/workflows 中对应任务的工作流即可开始。视频示例默认使用 33 帧、16 fps 和长边 640，约 2.06 秒，方便 24 GB 显存机器测试。

实测情况

测试设备：RTX 5090 Laptop 24 GB、64 GB 内存。
当前 Core 候选实现按官方 T2V 预设完成 640×368、33 帧测试。
无其他 GPU 计算进程的冷启动复测按官方 50/1/50 步在 1,024.58 秒（约 17 分 05 秒）完成，峰值 ComfyUI 可见显存 16.59 GiB。
此前同一兼容栈的成功基线为 1,756.39 秒、23.26 GiB；PyTorch 2.7 + CUDA 12.8 eager/offload 路径会受宿主机内存和进程状态影响，四小时未完成不属于正常基线。
外部 NVFP4 双 renderer 已通过 640×368、33 帧、25/5/40 步画质门； CUDA 13 优化速度和独占显存对比仍待验证。
Q4_K_S GGUF 双 renderer 已通过同规格画质门，33 帧全部唯一；双文件 16.31 GiB，相比 Balanced INT8 renderer 对减少 38.9%，但有轻微脸部变软和腿/尾粘连，因此仍是实验档。
T2V、V2V、R2V、RV2V 的两秒长边 640 测试均通过。
同一 ComfyUI 进程连续运行两个任务后，显存和内存可以回落。

详细数据见 docs/QUALITY_TESTS.md、 docs/SMOKE_TESTS.md 和 docs/LOW_MEMORY_WEIGHTS.md。

原生 Core 实现已提交到 ComfyUI PR #16019。旧的分片方案 #16001 已关闭。当前 PR 可合并、全部 CI 通过、无未解决审查线程，正在等待维护者审查。

自己转换权重

如果不想下载重打包权重，可以从官方 ByteDance/Bernini-Diffusers-v2 自行转换：

python tools/download_model.py --output models/ByteDance/Bernini-Diffusers-v2
python tools/repack_diffusers.py `
  --source models/ByteDance/Bernini-Diffusers-v2 `
  --output C:/path/to/ComfyUI/models/bernini_v2/Bernini-v2-bf16-native `
  --storage-dtype bfloat16
python tools/quantize_repack.py `
  --source C:/path/to/ComfyUI/models/bernini_v2/Bernini-v2-bf16-native `
  --output C:/path/to/ComfyUI/models/bernini_v2/Bernini-v2-balanced-int8 `
  --profile balanced --device cuda
python tools/export_single_files.py `
  --source C:/path/to/ComfyUI/models/bernini_v2/Bernini-v2-balanced-int8 `
  --output C:/path/to/Bernini-v2-single-int8 `
  --profile int8

前两步产生仅供转换使用的分片工作区；最后一步才生成用户安装的四个单体文件。 validate_single_files.py 会检查单文件合同、键、dtype、量化标记、tokenizer 和双专家一致性。 GGUF 的分片转换、量化和验证命令见 docs/LOW_MEMORY_WEIGHTS.md。

English

What this project does

Bernini v2 plans a scene and its motion with Qwen2.5-VL, then renders it with two Wan2.2 experts. This repository implements that pipeline with native ComfyUI model loading, memory management, conditioning and sampling APIs. It does not run a hidden Diffusers pipeline inside a node.

It supports all six released tasks: t2i, i2i, t2v, v2v, r2v, and rv2v.

Install

Install Bernini v2 (Native) from ComfyUI-Manager, or clone this repository into ComfyUI/custom_nodes, then restart ComfyUI.

Registry release 0.3.3 is reviewed and Active.

Download the recommended 45.63 GiB Balanced INT8 standalone files directly into the standard ComfyUI model folders:

hf download t8star/Bernini-V2-Comfy `
  --include "text_encoders/*_int8.safetensors" "diffusion_models/*_int8.safetensors" "vae/wan_2.1_vae.safetensors" `
  --local-dir C:/path/to/ComfyUI/models

The 83.04 GiB BF16 files use the _bf16.safetensors suffix in the same text_encoders/ and diffusion_models/ folders. Each component is one file; there is no Diffusers directory, shard index, or repack manifest. The same repository includes the verified Wan 2.1 VAE companion file:

hf download t8star/Bernini-V2-Comfy `
  --include "vae/wan_2.1_vae.safetensors" `
  --local-dir C:/path/to/ComfyUI/models

Alternatively run python tools/download_vae.py --output C:/path/to/ComfyUI/models/vae.

An experimental 16.31 GiB Q4_K_S renderer pair is also published. Install ComfyUI-GGUF, then download it to the diffusion-model directory:

hf download t8star/Bernini-V2-Comfy `
  --include "Bernini-v2-GGUF-Q4_K_S/*" `
  --local-dir C:/path/to/ComfyUI/models/diffusion_models

Open a workflow from examples/workflows. Video examples use 33 frames at 16 fps with a 640-pixel long edge so the complete pipeline can be tested on a 24 GB GPU.

On the RTX 5090 Laptop 24 GB test host, a clean reboot with no competing GPU process completed the Core-only official 50/1/50 T2V preset at 640x368 and 33 frames in 1,024.58 seconds (about 17 minutes 5 seconds), with 16.59 GiB peak ComfyUI-visible VRAM. An earlier successful run on the same PyTorch 2.7 + CUDA 12.8 eager/offload compatibility stack took 1,756.39 seconds and peaked at 23.26 GiB, so host memory/process state materially affects this legacy path.

Development and Core work

The implementation is covered by unit, workflow, real-weight quality, memory, and repeat-run tests. The native Core implementation is submitted as ComfyUI PR #16019, closing issue #15702. Review notes live in docs/CORE_MERGE_PLAN.md. The PR is currently mergeable with all CI checks passing and no unresolved review threads; it is awaiting maintainer review.

An external NVFP4 renderer pair also passes the production-step 640x368, 33-frame T2V visual gate. This was a CUDA 12.8 eager-path quality run; CUDA 13 optimized speed and isolated-memory results remain pending.

The Q4_K_S GGUF pair passes the same 640x368, 33-frame production-step gate. All frames are unique and motion remains coherent, with mild facial softness and occasional leg/tail merging versus Balanced INT8. It is an experimental low-storage lane rather than the default quality recommendation.

python -m pytest
python -m ruff check .
python -m ruff format --check .

License and credits

This project and the repackaged weights use the Apache License 2.0. Bernini v2 was created and released by ByteDance; see the official model, source code, and THIRD_PARTY_NOTICES.md.

Bernini v2 (Native)

Nodes (7)

Bernini v2 for ComfyUI

中文

这是什么