DeepSeek简介

今天文章主题将介绍当前爆火的LLM大模型DeepSeek。关于DeepSeek的成本和开源等优势网上已有大堆资料，在本文将不再赘述。本文将使用DeepSeek超强的编程能力实现ComfyUI插件。

重要：DeepSeek官网会员经常服务器忙中断服务，建议使用硅基提供的满血版服务（可在线网页和API使用，注册即送14元可玩很长时间）：https://cloud.siliconflow.cn/i/UiJMuo4o

笔者演示需求：

全程以DeepSeek自动化编程为主，人工引导LLM编码和修正错误。
以同为DeepSeek开源的Janus-Pro模型为例，实现Janus-Pro模型的图像反推和文生图功能的ComfyUI封装。并要求插件支持更多用户自定义配置，将模型的配置参数释放给用户灵活配置。

DeepSeek Janus-Pro封装

01. 初始需求

LLM需求实现好坏依赖输入信息详细和准确度。首先告诉LLM实现的业务需求和功能要求（最好有需求任务Task拆分设计），同时笔者还提供了一份Janus-Pro官方推理示例代码（尽量保证LLM编码正确性）。

请根据ComfyUI插件规范写一个Janus-Pro图像多模态理解反推提示词的插件。要求：1：独立Janus-Pro模型加载便于后续其他功能使用。2：Janus-Pro模型支持1B、7B、32B多种选项。模型加载参数设置为对应插件可配置参数并附带推荐默认值。3：提供4、8、全精度模型加载选项参数。Janus-Pro的使用如下所示：```
model_path = "deepseek-ai/Janus-Pro-7B"
更多演示代码见：https://github.com/deepseek-ai/Janus```

02. 代码重构和修正

然后模型经过深度思考输入对应项目代码。快速阅读理解后，如果不满足业务需求或新增任务，则继续与LLM对话修正输出结果。例如：

请根据上述代码修改：1：将模型选项改为仅deepseek-ai/Janus-Pro-1B、deepseek-ai/Janus-Pro-7B模型加载。2：利用vl_chat_processor.tokenizer消除JanusProLoader的tokenizer输出。3：precision支持4位、8位的精度加载。4：去掉trust_remote_code参数，将其内置到代码中。5：JanusProLoader 改为JanusProModelLoader，JanusProPrompt改为JanusProImageUnderstanding 6：将JanusProLoader和JanusProPrompt拆分成2个独立文件。7 :load_pil_images 使用如下代码：```
更多演示代码见：https://github.com/deepseek-ai/Janus```

03.变更需求

请将JanusProModelLoader改为本地目录替代cache_dir。首先检查本地目录models/Janus-Pro目录是否有对应模型，如果有则本地加载。如果没有则使用snapshot_download下载到本地models/Janus-Pro目录

ComfyUI插件演示

经过多轮对话，任务Task增加、修正建议引导、调试报错信息的输入，最终使用DeepSeek全自动编程完成插件封装，展示地址为：https://github.com/greengerong/ComfyUI-JanusPro-PL

JanusPro模型ComfyUI体验

现在就可以通过ComfyUI-Manager 的Git安装插件DeepSeek 实现的JanusPro插件体验和测试了。模型文末获取！

• • ComfyUI-JanusPro-PL插件：https://github.com/greengerong/ComfyUI-JanusPro-PL
• • 模型下载：在首次运行时模型会自动下载到本地目录 /ComfyUI/models/Janus-Pro/ 下。也可以手动放置该目录 模型文末网盘获取！

/ComfyUI/models/Janus-Pro/
├── Janus-Pro-7B/
│   ├── config.json
│   ├── pytorch_model.bin
│   └── ...
└── Janus-Pro-1B/
    └── ...

Flux文生图&混元视频工作流

最新LIBLIBAI平台已支持Flux文生图和混元视频ComfyUI工作流在线体验：

• • F.1-绮梦流光-水湄凝香：https://www.liblib.art/modelinfo/134c6dd95aef48e98a22b24e003e026b
• • 文生图-Flux文生图(PuLID|LORA|Joy|SUPIR)工作流：https://www.liblib.art/modelinfo/782aacd70f604da39e83368c696a02a8?versionUuid=9c5eceb01fb94d4d93d60fe2c0bd7468
• • 文生视频-腾迅混元最强开源视频(LORA)工作流：https://www.liblib.art/modelinfo/35ee21d5f6a94204abb767ad194ab9cd?versionUuid=be674032ffa14e5597a08922556f4da0

JanusPro模型ComfyUI工作流体验

JanusPro模型ComfyUI工作流已上传LIBLIBAI平台可体验：https://www.liblib.art/modelinfo/73989dd3503d4b76a66d9fcac88c6d45?versionUuid=4e3c4bb2eee243498164da9713d94542

注意：

• • JanusPro模型的文生图当前仅支持384分辨率，且仅为文生图雏形质量不高，本文将不做演示使用Flux模型替代。
• • 插件支持多精度和1B/7B模型选项，对于低显存可选用1B模型和int4运行预估6G可运行推理。当前仅做了功能验证，未进一步性能验证和优化，如遇见问题github issue反馈，将持续优化改进。
• • 以下案例使用Flux-绮梦流光-极简水墨国风_极简水墨LORA模型：https://www.liblib.art/modelinfo/99d6959cc50c46bd974d1012bf60f5de?versionUuid=f984e20cf18241c2a5e95f09b907b5a6

01.旅行

pl,Chinese ink painting, The image depicts a man dressed in traditional attire, standing on an airport tarmac. He is wearing a green robe with a red belt and has long black hair tied back in a ponytail. He is holding a large gray suitcase with wheels in his right hand and appears to be gesturing with his left hand. In the background, there is a commercial airplane with a blue and white color scheme. The airplane is stationary on the runway, and the scene suggests that the man is either about to board the plane or has just disembarked. The overall style of the image is a blend of traditional and modern elements, creating a unique visual contrast.

02.熊猫

pl,Chinese ink painting, This image depicts a serene and whimsical scene featuring a panda lounging in a hammock. The panda is comfortably positioned, with its body relaxed and its head slightly tilted back, giving an impression of contentment and tranquility. The hammock is suspended between two bamboo poles, which are part of a bamboo grove in the background. The bamboo stalks are tall and slender, with leaves extending upwards and outwards, creating a natural canopy. The color palette is predominantly monochromatic, with shades of black, white, and gray, giving the image a calm and soothing atmosphere. The ground beneath the hammock is lightly shaded, and there are a few stacked books or boxes near the base of the bamboo poles, adding a touch of human presence to the otherwise natural setting. The overall composition exudes a sense of peace and relaxation, inviting the viewer to imagine themselves in this serene environment.

03.舞蹈

pl,Chinese ink painting, The image depicts a group of ballet dancers in mid-performance. The central figure is a female dancer wearing a vibrant red dress with a deep V-neckline and a flared skirt. The dress has a black trim at the hem and waist, adding contrast to the bright red fabric. She is captured in a graceful leap, with one leg extended behind her and her arms elegantly raised, forming a perfect arabesque. To her left, another female dancer is also in mid-air, wearing a gray dress with a similar V-neckline and flared skirt. This dancer's dress also features a black trim, and her pose mirrors that of the central dancer, with one leg extended and arms gracefully raised. In the background, there are additional dancers, slightly faded, suggesting depth and movement. They are also dressed in ballet attire, with gray and white dresses, and their poses indicate they are engaged in a synchronized dance routine. The overall composition of the image is dynamic, capturing the elegance and precision of ballet. The use of watercolor-like strokes and soft shading adds a sense of fluidity and grace to the scene.

这里为了帮助大家更好地掌握 ComfyUI，分享一套字节大佬整理的ComfyUI工作流集合，其包含了很多好玩有趣，但又有点复杂的工作流节点和json配置。

涵盖了 Stable Scascade、3D、LLM+SD、Portrait Master、SVD 等相关类别的工作流，共计15个类目38项工作流。这些都放在了下方卡片，需要的点击免费获取：

对于初学者来说，最佳的学习方法是以这些现成的工作流为模板，一步步地复刻并理解它们。

通过观察和分析别人的工作流，你可以学习到各种节点搭建的技巧和方法。随着理解的深入，你将能够根据自己的需求创新和搭建属于自己的工作流。

希望本文能帮助你有效地提升你的设计效率和创造力。

对于从来没有接触过AI绘画的同学，我已经帮你们准备了详细的学习成长路线图。可以说是最科学最系统的学习路线，大家跟着这个大的方向学习准没问题。

这份完整版的AI绘画资料和SD整合包已经打包好了，需要的点击下方插件，即可前往免费领取！