模型资源

简介

模型分类

graph TD
    A[模型资源] --> B[基础模型]
    A --> C[LoRA模型]
    A --> D[VAE模型]
    A --> E[ControlNet模型]
    A --> F[专用模型]

    B --> B1[SD1.5系列]
    B --> B2[SD2.1系列]
    B --> B3[SDXL系列]

    C --> C1[风格LoRA]
    C --> C2[人物LoRA]
    C --> C3[概念LoRA]

    D --> D1[标准VAE]
    D --> D2[优化VAE]

    E --> E1[边缘检测]
    E --> E2[姿态估计]
    E --> E3[深度估计]

    F --> F1[动漫模型]
    F --> F2[写实模型]
    F --> F3[艺术模型]

    style A fill:#e1f5ff
    style B fill:#e1ffe1
    style C fill:#e1ffe1
    style D fill:#e1ffe1
    style E fill:#e1ffe1
    style F fill:#e1ffe1

基础模型

1. Stable Diffusion 1.5系列

Realistic Vision

{
  "model": "realisticVisionV51_v51VAE.safetensors",
  "best_for": "人像、写实风格",
  "resolution": "512x512",
  "recommended_steps": 30,
  "recommended_cfg": 7.0,
  "tips": "适合生成高质量人像，配合人像LoRA效果更佳"
}

DreamShaper

{
  "model": "dreamshaper_8.safetensors",
  "best_for": "艺术创作、创意设计",
  "resolution": "512x512",
  "recommended_steps": 25,
  "recommended_cfg": 7.5,
  "tips": "适合艺术风格创作，色彩表现丰富"
}

2. Stable Diffusion 2.1系列

SD 2.1 768

{
  "model": "v2-1_768-ema-pruned.safetensors",
  "best_for": "高分辨率图像",
  "resolution": "768x768",
  "recommended_steps": 35,
  "recommended_cfg": 7.0,
  "tips": "支持更高分辨率，适合需要细节的场景"
}

3. Stable Diffusion XL系列

SDXL Base 1.0

{
  "model": "sd_xl_base_1.0.safetensors",
  "best_for": "最高质量图像",
  "resolution": "1024x1024",
  "recommended_steps": 50,
  "recommended_cfg": 7.5,
  "tips": "当前最高质量的模型，配合Refiner使用效果更佳"
}

SDXL Turbo

{
  "model": "sdxl_turbo.safetensors",
  "best_for": "快速预览",
  "resolution": "512x512",
  "recommended_steps": 4-10,
  "recommended_cfg": 1.0-2.0,
  "tips": "极快生成速度，适合快速预览和迭代"
}

LoRA模型

4. 风格LoRA

Detail Tweaker

{
  "lora": "detail_tweaker.safetensors",
  "strength": 0.7,
  "best_for": "增强图像细节",
  "compatible_models": ["SD1.5", "SD2.1"],
  "tips": "适度使用，过度使用可能导致过度锐化"
}

5. 人物LoRA

Epic Photogasm

{
  "lora": "epic_photogasm.safetensors",
  "strength": 0.6,
  "best_for": "人像摄影",
  "compatible_models": ["Realistic Vision", "Deliberate"],
  "tips": "配合写实模型使用，效果最佳"
}

6. 概念LoRA

Cyberpunk Style

{
  "lora": "cyberpunk_style.safetensors",
  "strength": 0.7,
  "best_for": "科幻、未来主义",
  "compatible_models": ["SD1.5", "DreamShaper"],
  "tips": "适合创作科幻题材作品"
}

VAE模型

7. 标准VAE

VAE-ft-MSE-840000

{
  "vae": "vae-ft-mse-840000.safetensors",
  "best_for": "SD1.5系列模型",
  "benefits": "色彩还原准确，细节表现好",
  "tips": "SD1.5模型的默认推荐VAE"
}

8. 优化VAE

ControlNet模型

9. 边缘检测

Canny

{
  "controlnet": "control_v11p_sd15_canny.safetensors",
  "best_for": "轮廓控制",
  "recommended_strength": 0.8-1.0,
  "tips": "适合需要精确控制轮廓的场景"
}

10. 姿态估计

OpenPose

{
  "controlnet": "control_v11p_sd15_openpose.safetensors",
  "best_for": "人物姿态控制",
  "recommended_strength": 0.7-1.0,
  "tips": "适合人物创作，可以精确控制姿态"
}

11. 深度估计

Depth

{
  "controlnet": "control_v11f1p_sd15_depth.safetensors",
  "best_for": "空间深度控制",
  "recommended_strength": 0.5-0.8,
  "tips": "适合需要控制空间关系的场景"
}

专用模型

12. 动漫模型

13. 写实模型

14. 艺术模型

模型下载

下载源

{
  "official_sources": {
    "HuggingFace": {
      "url": "https://huggingface.co",
      "description": "官方模型托管平台",
      "reliability": "高",
      "speed": "快"
    },
    "Stability AI": {
      "url": "https://stability.ai",
      "description": "Stability AI官方",
      "reliability": "高",
      "speed": "中"
    }
  }
}

{
  "community_sources": {
    "CivitAI": {
      "url": "https://civitai.com",
      "description": "最大的AI艺术社区",
      "reliability": "高",
      "speed": "快"
    },
    "LiblibAI": {
      "url": "https://www.liblib.art",
      "description": "中文AI艺术社区",
      "reliability": "高",
      "speed": "快"
    }
  }
}

下载工具

# 使用huggingface-cli下载
import subprocess

def download_from_huggingface(model_id, output_dir):
    """从HuggingFace下载模型"""
    cmd = [
        "huggingface-cli",
        "download",
        model_id,
        "--local-dir",
        output_dir
    ]
    subprocess.run(cmd)

# 使用示例
# download_from_huggingface("runwayml/stable-diffusion-v1-5", "./models")

模型管理

模型组织

models/
├── checkpoints/          # 基础模型
│   ├── SD1.5/
│   ├── SD2.1/
│   └── SDXL/
├── lora/                # LoRA模型
│   ├── style/
│   ├── character/
│   └── concept/
├── vae/                 # VAE模型
├── controlnet/          # ControlNet模型
└── embeddings/         # 文本嵌入

模型验证

import torch
from safetensors import safe_open

def validate_model(model_path):
    """验证模型文件"""
    try:
        # 检查文件是否存在
        if not os.path.exists(model_path):
            return False, "文件不存在"

        # 检查文件大小
        file_size = os.path.getsize(model_path)
        if file_size < 1024 * 1024:  # 小于1MB
            return False, "文件过小"

        # 尝试加载模型
        if model_path.endswith('.safetensors'):
            with safe_open(model_path, framework="pt", device="cpu") as f:
                keys = f.keys()
                if len(keys) == 0:
                    return False, "模型为空"

        return True, "验证通过"

    except Exception as e:
        return False, str(e)

模型选择指南

选择流程

graph TD
    A[选择模型] --> B{需求类型?}

    B -->|高质量| C[SDXL]
    B -->|快速| D[SD-Turbo]
    B -->|人像| E[Realistic Vision]
    B -->|动漫| F[Anything V5]
    B -->|艺术| G[DreamShaper]
    B -->|通用| H[SD1.5]

    C --> C1[1024x1024]
    D --> D2[512x512]
    E --> E3[512x512]
    F --> F4[512x512]
    G --> G5[512x512]
    H --> H6[512x512]

    style A fill:#e1f5ff
    style C fill:#e1ffe1
    style D fill:#e1ffe1
    style E fill:#e1ffe1
    style F fill:#e1ffe1
    style G fill:#e1ffe1
    style H fill:#e1ffe1

选择建议

{
  "usage_based_selection": {
    "portrait": {
      "primary": "Realistic Vision",
      "secondary": ["Deliberate", "Majicmix Realistic"],
      "lora": ["Epic Photogasm", "Portrait Helper"]
    },
    "landscape": {
      "primary": "DreamShaper",
      "secondary": ["SD1.5", "SDXL"],
      "lora": ["Add More Details", "Detail Tweaker"]
    },
    "anime": {
      "primary": "Anything V5",
      "secondary": ["Counterfeit V3", "GhostMix"],
      "lora": ["Anime Style", "Character LoRA"]
    },
    "artistic": {
      "primary": "DreamShaper",
      "secondary": ["Protogen", "Rev Animated"],
      "lora": ["Style LoRA", "Concept LoRA"]
    }
  }
}

最佳实践

模型使用建议

1. 从基础开始: 先使用基础模型熟悉
2. 逐步添加: 逐步添加LoRA和ControlNet
3. 测试效果: 测试不同组合的效果
4. 记录配置: 记录有效的配置组合
5. 定期更新: 关注新模型发布

性能优化

1. 选择合适模型: 根据硬件选择模型
2. 使用量化模型: 在低显存设备上使用量化模型
3. 优化加载: 预加载常用模型
4. 缓存结果: 缓存中间结果

总结

模型资源是ComfyUI创作的核心。关键要点：

1. 了解模型特点: 了解不同模型的特点和适用场景
2. 合理选择: 根据需求选择合适的模型
3. 组合使用: 合理组合使用多个模型
4. 持续学习: 关注新模型和新技术

通过合理使用模型资源，可以创作出高质量的图像作品。