|
| 1 | +--- |
| 2 | +title: "Kandinsky 5.0" |
| 3 | +description: "本指南介绍如何在 ComfyUI 中使用 Kandinsky 5.0 视频生成工作流" |
| 4 | +sidebarTitle: "Kandinsky 5.0" |
| 5 | +--- |
| 6 | + |
| 7 | +import UpdateReminder from "/snippets/zh/tutorials/update-reminder.mdx"; |
| 8 | + |
| 9 | +[Kandinsky 5.0](https://huggingface.co/kandinskylab/Kandinsky-5.0-I2V-Lite-5s) 是由 [Kandinsky Lab](https://huggingface.co/kandinskylab) 开发的视频和图像生成扩散模型系列。Kandinsky 5.0 T2V Lite 是一个轻量级的 2B 参数模型,在开源视频生成模型中名列前茅,能够生成长达 10 秒的视频。 |
| 10 | + |
| 11 | +<UpdateReminder/> |
| 12 | + |
| 13 | +## 概述 |
| 14 | + |
| 15 | +Kandinsky 5.0 使用带有 Flow Matching 的潜在扩散管道,具有以下特点: |
| 16 | + |
| 17 | +- **扩散 Transformer (DiT):** 主要生成骨干网络,通过交叉注意力连接文本嵌入 |
| 18 | +- **Qwen2.5-VL 和 CLIP:** 提供高质量的文本嵌入 |
| 19 | +- **HunyuanVideo 3D VAE:** 将视频编码和解码到潜在空间 |
| 20 | + |
| 21 | +该模型系列包含多个针对不同用例优化的变体: |
| 22 | +- **SFT 模型:** 最高生成质量 |
| 23 | +- **CFG-distilled:** 推理速度提升 2 倍 |
| 24 | +- **Diffusion-distilled:** 速度提升 6 倍,质量损失极小(16 步) |
| 25 | +- **Pretrain 模型:** 专为微调设计 |
| 26 | + |
| 27 | +所有模型均提供 5 秒和 10 秒视频生成版本。 |
| 28 | + |
| 29 | +## 模型变体 |
| 30 | + |
| 31 | +| 模型 | 视频时长 | NFE | 延迟 (H100) | |
| 32 | +|-------|---------------|-----|----------------| |
| 33 | +| Kandinsky 5.0 T2V Lite SFT | 5s / 10s | 100 | 139s / 224s | |
| 34 | +| Kandinsky 5.0 T2V Lite no-CFG | 5s / 10s | 50 | 77s / 124s | |
| 35 | +| Kandinsky 5.0 T2V Lite distill | 5s / 10s | 16 | 35s / 61s | |
| 36 | +| Kandinsky 5.0 I2V Lite | 5s | 100 | 673s | |
| 37 | + |
| 38 | +## 文生视频工作流 |
| 39 | + |
| 40 | +### 1. 下载工作流文件 |
| 41 | + |
| 42 | +请更新你的 ComfyUI 到最新版本,并通过菜单 `工作流` -> `浏览模板` -> `视频` 找到 "Kandinsky 5.0 T2V" 以加载工作流。 |
| 43 | + |
| 44 | +<a className="prose" target='_blank' href="https://raw.githubusercontent.com/Comfy-Org/workflow_templates/refs/heads/main/templates/video_kandinsky5_t2v.json" style={{ display: 'inline-block', backgroundColor: '#0078D6', color: '#ffffff', padding: '10px 20px', borderRadius: '8px', borderColor: "transparent", textDecoration: 'none', fontWeight: 'bold'}}> |
| 45 | + <p className="prose" style={{ margin: 0, fontSize: "0.8rem" }}>下载 JSON 格式工作流</p> |
| 46 | +</a> |
| 47 | + |
| 48 | +### 2. 手动下载模型 |
| 49 | + |
| 50 | +**Text Encoders** |
| 51 | +- [qwen_2.5_vl_7b_fp8_scaled.safetensors](https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors) |
| 52 | +- [clip_l.safetensors](https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/clip_l.safetensors) |
| 53 | + |
| 54 | +**Diffusion Model** |
| 55 | +- [kandinsky5lite_t2v_sft_5s.safetensors](https://huggingface.co/kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s/resolve/main/model/kandinsky5lite_t2v_sft_5s.safetensors) |
| 56 | + |
| 57 | +**VAE** |
| 58 | +- [hunyuan_video_vae_bf16.safetensors](https://huggingface.co/Kijai/HunyuanVideo_comfy/resolve/main/hunyuan_video_vae_bf16.safetensors) |
| 59 | + |
| 60 | +``` |
| 61 | +ComfyUI/ |
| 62 | +├── 📂 models/ |
| 63 | +│ ├── 📂 text_encoders/ |
| 64 | +│ │ ├── qwen_2.5_vl_7b_fp8_scaled.safetensors |
| 65 | +│ │ └── clip_l.safetensors |
| 66 | +│ ├── 📂 diffusion_models/ |
| 67 | +│ │ └── kandinsky5lite_t2v_sft_5s.safetensors |
| 68 | +│ └── 📂 vae/ |
| 69 | +│ └── hunyuan_video_vae_bf16.safetensors |
| 70 | +``` |
| 71 | + |
| 72 | +## 图生视频工作流 |
| 73 | + |
| 74 | +### 1. 下载工作流文件 |
| 75 | + |
| 76 | +请更新你的 ComfyUI 到最新版本,并通过菜单 `工作流` -> `浏览模板` -> `视频` 找到 "Kandinsky 5.0 I2V" 以加载工作流。 |
| 77 | + |
| 78 | +<a className="prose" target='_blank' href="https://raw.githubusercontent.com/Comfy-Org/workflow_templates/refs/heads/main/templates/video_kandinsky5_i2v.json" style={{ display: 'inline-block', backgroundColor: '#0078D6', color: '#ffffff', padding: '10px 20px', borderRadius: '8px', borderColor: "transparent", textDecoration: 'none', fontWeight: 'bold'}}> |
| 79 | + <p className="prose" style={{ margin: 0, fontSize: "0.8rem" }}>下载 JSON 格式工作流</p> |
| 80 | +</a> |
| 81 | + |
| 82 | +### 2. 手动下载模型 |
| 83 | + |
| 84 | +**Text Encoders** |
| 85 | +- [qwen_2.5_vl_7b_fp8_scaled.safetensors](https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors) |
| 86 | +- [clip_l.safetensors](https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/clip_l.safetensors) |
| 87 | + |
| 88 | +**Diffusion Model** |
| 89 | +- [kandinsky5lite_i2v_sft_5s.safetensors](https://huggingface.co/kandinskylab/Kandinsky-5.0-I2V-Lite-5s/resolve/main/model/kandinsky5lite_i2v_sft_5s.safetensors) |
| 90 | + |
| 91 | +**VAE** |
| 92 | +- [hunyuan_video_vae_bf16.safetensors](https://huggingface.co/Kijai/HunyuanVideo_comfy/resolve/main/hunyuan_video_vae_bf16.safetensors) |
| 93 | + |
| 94 | +``` |
| 95 | +ComfyUI/ |
| 96 | +├── 📂 models/ |
| 97 | +│ ├── 📂 text_encoders/ |
| 98 | +│ │ ├── qwen_2.5_vl_7b_fp8_scaled.safetensors |
| 99 | +│ │ └── clip_l.safetensors |
| 100 | +│ ├── 📂 diffusion_models/ |
| 101 | +│ │ └── kandinsky5lite_i2v_sft_5s.safetensors |
| 102 | +│ └── 📂 vae/ |
| 103 | +│ └── hunyuan_video_vae_bf16.safetensors |
| 104 | +``` |
| 105 | + |
| 106 | +## 资源 |
| 107 | + |
| 108 | +- [HuggingFace 模型合集](https://huggingface.co/collections/kandinskylab/kandinsky-50-video-lite) |
| 109 | +- [GitHub 仓库](https://github.com/ai-forever/Kandinsky-5) |
| 110 | +- [ComfyUI 集成](https://github.com/ai-forever/Kandinsky-5/blob/main/comfyui/README.md) |
| 111 | +- [项目主页](https://ai-forever.github.io/Kandinsky-5/) |
0 commit comments