[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-Xiaojiu-z--EasyControl":3,"tool-Xiaojiu-z--EasyControl":64},[4,17,27,35,43,56],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":16},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,3,"2026-04-05T11:01:52",[13,14,15],"开发框架","图像","Agent","ready",{"id":18,"name":19,"github_repo":20,"description_zh":21,"stars":22,"difficulty_score":23,"last_commit_at":24,"category_tags":25,"status":16},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",140436,2,"2026-04-05T23:32:43",[13,15,26],"语言模型",{"id":28,"name":29,"github_repo":30,"description_zh":31,"stars":32,"difficulty_score":23,"last_commit_at":33,"category_tags":34,"status":16},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",107662,"2026-04-03T11:11:01",[13,14,15],{"id":36,"name":37,"github_repo":38,"description_zh":39,"stars":40,"difficulty_score":23,"last_commit_at":41,"category_tags":42,"status":16},3704,"NextChat","ChatGPTNextWeb\u002FNextChat","NextChat 是一款轻量且极速的 AI 助手，旨在为用户提供流畅、跨平台的大模型交互体验。它完美解决了用户在多设备间切换时难以保持对话连续性，以及面对众多 AI 模型不知如何统一管理的痛点。无论是日常办公、学习辅助还是创意激发，NextChat 都能让用户随时随地通过网页、iOS、Android、Windows、MacOS 或 Linux 端无缝接入智能服务。\n\n这款工具非常适合普通用户、学生、职场人士以及需要私有化部署的企业团队使用。对于开发者而言，它也提供了便捷的自托管方案，支持一键部署到 Vercel 或 Zeabur 等平台。\n\nNextChat 的核心亮点在于其广泛的模型兼容性，原生支持 Claude、DeepSeek、GPT-4 及 Gemini Pro 等主流大模型，让用户在一个界面即可自由切换不同 AI 能力。此外，它还率先支持 MCP（Model Context Protocol）协议，增强了上下文处理能力。针对企业用户，NextChat 提供专业版解决方案，具备品牌定制、细粒度权限控制、内部知识库整合及安全审计等功能，满足公司对数据隐私和个性化管理的高标准要求。",87618,"2026-04-05T07:20:52",[13,26],{"id":44,"name":45,"github_repo":46,"description_zh":47,"stars":48,"difficulty_score":23,"last_commit_at":49,"category_tags":50,"status":16},2268,"ML-For-Beginners","microsoft\u002FML-For-Beginners","ML-For-Beginners 是由微软推出的一套系统化机器学习入门课程，旨在帮助零基础用户轻松掌握经典机器学习知识。这套课程将学习路径规划为 12 周，包含 26 节精炼课程和 52 道配套测验，内容涵盖从基础概念到实际应用的完整流程，有效解决了初学者面对庞大知识体系时无从下手、缺乏结构化指导的痛点。\n\n无论是希望转型的开发者、需要补充算法背景的研究人员，还是对人工智能充满好奇的普通爱好者，都能从中受益。课程不仅提供了清晰的理论讲解，还强调动手实践，让用户在循序渐进中建立扎实的技能基础。其独特的亮点在于强大的多语言支持，通过自动化机制提供了包括简体中文在内的 50 多种语言版本，极大地降低了全球不同背景用户的学习门槛。此外，项目采用开源协作模式，社区活跃且内容持续更新，确保学习者能获取前沿且准确的技术资讯。如果你正寻找一条清晰、友好且专业的机器学习入门之路，ML-For-Beginners 将是理想的起点。",84991,"2026-04-05T10:45:23",[14,51,52,53,15,54,26,13,55],"数据工具","视频","插件","其他","音频",{"id":57,"name":58,"github_repo":59,"description_zh":60,"stars":61,"difficulty_score":10,"last_commit_at":62,"category_tags":63,"status":16},3128,"ragflow","infiniflow\u002Fragflow","RAGFlow 是一款领先的开源检索增强生成（RAG）引擎，旨在为大语言模型构建更精准、可靠的上下文层。它巧妙地将前沿的 RAG 技术与智能体（Agent）能力相结合，不仅支持从各类文档中高效提取知识，还能让模型基于这些知识进行逻辑推理和任务执行。\n\n在大模型应用中，幻觉问题和知识滞后是常见痛点。RAGFlow 通过深度解析复杂文档结构（如表格、图表及混合排版），显著提升了信息检索的准确度，从而有效减少模型“胡编乱造”的现象，确保回答既有据可依又具备时效性。其内置的智能体机制更进一步，使系统不仅能回答问题，还能自主规划步骤解决复杂问题。\n\n这款工具特别适合开发者、企业技术团队以及 AI 研究人员使用。无论是希望快速搭建私有知识库问答系统，还是致力于探索大模型在垂直领域落地的创新者，都能从中受益。RAGFlow 提供了可视化的工作流编排界面和灵活的 API 接口，既降低了非算法背景用户的上手门槛，也满足了专业开发者对系统深度定制的需求。作为基于 Apache 2.0 协议开源的项目，它正成为连接通用大模型与行业专有知识之间的重要桥梁。",77062,"2026-04-04T04:44:48",[15,14,13,26,54],{"id":65,"github_repo":66,"name":67,"description_en":68,"description_zh":69,"ai_summary_zh":69,"readme_en":70,"readme_zh":71,"quickstart_zh":72,"use_case_zh":73,"hero_image_url":74,"owner_login":75,"owner_name":76,"owner_avatar_url":77,"owner_bio":76,"owner_company":78,"owner_location":79,"owner_email":76,"owner_twitter":76,"owner_website":76,"owner_url":80,"languages":81,"stars":94,"forks":95,"last_commit_at":96,"license":97,"difficulty_score":10,"env_os":98,"env_gpu":99,"env_ram":100,"env_deps":101,"category_tags":112,"github_topics":76,"view_count":23,"oss_zip_url":76,"oss_zip_packed_at":76,"status":16,"created_at":113,"updated_at":114,"faqs":115,"releases":151},2166,"Xiaojiu-z\u002FEasyControl","EasyControl","Implementation of \"EasyControl: Adding Efficient and Flexible Control for Diffusion Transformer\"(ICCV2025)","EasyControl 是一款专为新一代扩散变压器（DiT）架构设计的开源控制框架，旨在为图像生成模型提供高效且灵活的精准控制能力。随着 AI 绘图技术从传统的 Unet 架构向 DiT 架构演进，生态系统中长期缺乏成熟的插件支持，常面临计算效率低、多条件控制冲突以及模型适应性不足等痛点。EasyControl 通过引入轻量级的条件注入 LoRA 模块、位置感知训练范式，并结合因果注意力机制与 KV Cache 技术，成功解决了这些难题。\n\n它不仅实现了“即插即用”的便捷性，还能在保持风格无损的前提下，灵活支持多种分辨率、宽高比及复杂的多条件组合控制，显著提升了推理速度与生成质量。近期更新更集成了 CFG-Zero* 技术以增强画面保真度，并推出了独特的吉卜力风格迁移模型，仅需少量数据即可实现高质量的角色风格化。\n\n这款工具非常适合 AI 研究人员探索 DiT 架构潜力，开发者构建定制化绘图应用，以及设计师和高级爱好者进行高精度的创意创作。无论是需要微调模型的专业团队，还是希望在本地部署高效生成流程的技术用户，EasyControl 都提供了一个强大而友好的解决方案。","# Implementation of EasyControl\n\nEasyControl: Adding Efficient and Flexible Control for Diffusion Transformer\n\n\u003Ca href='https:\u002F\u002Feasycontrolproj.github.io\u002F'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject-Page-green'>\u003C\u002Fa>\n\u003Ca href='https:\u002F\u002Farxiv.org\u002Fpdf\u002F2503.07027'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FTechnique-Report-red'>\u003C\u002Fa> \n\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002FXiaojiu-Z\u002FEasyControl\u002F\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F🤗_HuggingFace-Model-ffbd45.svg\" alt=\"HuggingFace\">\u003C\u002Fa>\n\u003Ca href='https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fjamesliu1217\u002FEasyControl'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20Hugging%20Face-Spaces-blue'>\u003C\u002Fa>\n\u003Ca href='https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fjamesliu1217\u002FEasyControl_Ghibli'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20Hugging%20Face-Ghibli_Spaces-purple'>\u003C\u002Fa>\n\n> *[Yuxuan Zhang](https:\u002F\u002Fxiaojiu-z.github.io\u002FYuxuanZhang.github.io\u002F), [Yirui Yuan](https:\u002F\u002Fgithub.com\u002FReynoldyy), [Yiren Song](https:\u002F\u002Fscholar.google.com.hk\u002Fcitations?user=L2YS0jgAAAAJ), [Haofan Wang](https:\u002F\u002Fhaofanwang.github.io\u002F), [Jiaming Liu](https:\u002F\u002Fscholar.google.com\u002Fcitations?user=SmL7oMQAAAAJ&hl=en)*\n> \u003Cbr>\n> The Chinese University of Hong Kong, ShanghaiTech University, Tiamat AI, Alibaba Group, National University of Singapore, Liblib AI\n\n\u003Cimg src='https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_3a80e064e594.jpg'>\n\n## Features\n* **Motivation:** The architecture of diffusion models is transitioning from Unet-based to DiT (Diffusion Transformer). However, the DiT ecosystem lacks mature plugin support and faces challenges such as efficiency bottlenecks, conflicts in multi-condition coordination, and insufficient model adaptability.\n* **Contribution:** We propose EasyControl, an efficient and flexible unified conditional DiT framework. By incorporating a lightweight Condition Injection LoRA module, a Position-Aware Training Paradigm, and a combination of Causal Attention mechanisms with KV Cache technology, we significantly enhance **model compatibility** (enabling plug-and-play functionality and style lossless control), **generation flexibility** (supporting multiple resolutions, aspect ratios, and multi-condition combinations), and **inference efficiency**.\n\u003Cimg src='https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_c6d83e10bf5f.jpg'>\n\n## News\n- **2025-04-11**: 🔥🔥🔥 Training code have been released. Recommanded Hardware: at least 1x NVIDIA H100\u002FH800\u002FA100, GPUs Memory: ~80GB GPU memory.\n- **2025-04-09**: ⭐️ The codes for the simple API have been released. If you wish to run the models on your personal machines, head over to the simple_api branch to access the relevant resources.\n\n- **2025-04-07**: 🔥 Thanks to the great work by [CFG-Zero*](https:\u002F\u002Fgithub.com\u002FWeichenFan\u002FCFG-Zero-star) team, EasyControl is now integrated with CFG-Zero*!! With just a few lines of code, you can boost image fidelity and controllability!! You can download the modified code from [this link](https:\u002F\u002Fgithub.com\u002FWeichenFan\u002FCFG-Zero-star\u002Fblob\u002Fmain\u002Fmodels\u002Feasycontrol\u002Finfer.py) and try it.\n\n\u003Ctable class=\"center\">\n  \u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_d05c5684cb60.webp\" style=\"width:410px; height:auto;\">\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_22c4d532bd34.webp\" style=\"width:410px; height:auto;\">\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_2f830fe19776.webp\" style=\"width:410px; height:auto;\">\u003C\u002Ftd>\n  \u003C\u002Ftr>\n  \u003Ctr>\n    \u003Ctd align=\"center\">\u003Cb>Source Image\u003C\u002Fb>\u003C\u002Ftd>\n    \u003Ctd align=\"center\">\u003Cb>CFG\u003C\u002Fb>\u003C\u002Ftd>\n    \u003Ctd align=\"center\">\u003Cb>CFG-Zero*\u003C\u002Fb>\u003C\u002Ftd>\n  \u003C\u002Ftr>\n\u003C\u002Ftable>\n\n- **2025-04-03**: Thanks to [jax-explorer](https:\u002F\u002Fgithub.com\u002Fjax-explorer), [Ghibli Img2Img Control ComfyUI Node](https:\u002F\u002Fgithub.com\u002Fjax-explorer\u002FComfyUI-easycontrol) is supported!\n\n- **2025-04-01**: 🔥 New [Stylized Img2Img Control Model](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fjamesliu1217\u002FEasyControl_Ghibli) is now released!! Transform portraits into Studio Ghibli-style artwork using this LoRA model. Trained on **only 100 real Asian faces** paired with **GPT-4o-generated Ghibli-style counterparts**, it preserves facial features while applying the iconic anime aesthetic.\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_94779297bdc3.jpeg\" alt=\"Example 3\" width=\"400\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_0982b737852d.jpeg\" alt=\"Example 4\" width=\"400\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">Example 3\u003C\u002Ftd>\n    \u003Ctd align=\"center\">Example 4\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n- **2025-03-19**: 🔥 We have released [huggingface demo](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fjamesliu1217\u002FEasyControl)! You can now try out EasyControl with the huggingface space, enjoy it!\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_8bc7770e095f.jpeg\" alt=\"Example 1\" width=\"400\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_04e8860a7b68.jpeg\" alt=\"Example 2\" width=\"400\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">Example 1\u003C\u002Ftd>\n    \u003Ctd align=\"center\">Example 2\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n- **2025-03-18**: 🔥 We have released our [pre-trained checkpoints](https:\u002F\u002Fhuggingface.co\u002FXiaojiu-Z\u002FEasyControl\u002F) on Hugging Face! You can now try out EasyControl with the official weights. \n- **2025-03-12**: ⭐️ Inference code are released. Once we have ensured that everything is functioning correctly, the new model will be merged into this repository. Stay tuned for updates! 😊\n\n## Installation\n\nWe recommend using Python 3.10 and PyTorch with CUDA support. To set up the environment:\n\n```bash\n# Create a new conda environment\nconda create -n easycontrol python=3.10\nconda activate easycontrol\n\n# Install other dependencies\npip install -r requirements.txt\n```\n\n## Download\n\nYou can download the model directly from [Hugging Face](https:\u002F\u002Fhuggingface.co\u002FEasyControl\u002FEasyControl).\nOr download using Python script:\n\n```python\nfrom huggingface_hub import hf_hub_download\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fcanny.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fdepth.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fhedsketch.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Finpainting.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fpose.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fseg.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fsubject.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002FGhibli.safetensors\", local_dir=\".\u002F\")\n```\n\nIf you cannot access Hugging Face, you can use [hf-mirror](https:\u002F\u002Fhf-mirror.com\u002F) to download the models:\n```python\nexport HF_ENDPOINT=https:\u002F\u002Fhf-mirror.com\nhuggingface-cli download --resume-download Xiaojiu-Z\u002FEasyControl --local-dir checkpoints --local-dir-use-symlinks False\n```\n\n## Usage\nHere's a basic example of using EasyControl:\n\n### Model Initialization\n\n```python\nimport torch\nfrom PIL import Image\nfrom src.pipeline import FluxPipeline\nfrom src.transformer_flux import FluxTransformer2DModel\nfrom src.lora_helper import set_single_lora, set_multi_lora\n\ndef clear_cache(transformer):\n    for name, attn_processor in transformer.attn_processors.items():\n        attn_processor.bank_kv.clear()\n\n# Initialize model\ndevice = \"cuda\"\nbase_path = \"FLUX.1-dev\"  # Path to your base model\npipe = FluxPipeline.from_pretrained(base_path, torch_dtype=torch.bfloat16, device=device)\ntransformer = FluxTransformer2DModel.from_pretrained(\n    base_path, \n    subfolder=\"transformer\",\n    torch_dtype=torch.bfloat16, \n    device=device\n)\npipe.transformer = transformer\npipe.to(device)\n\n# Load control models\nlora_path = \".\u002Fcheckpoints\u002Fmodels\"\ncontrol_models = {\n    \"canny\": f\"{lora_path}\u002Fcanny.safetensors\",\n    \"depth\": f\"{lora_path}\u002Fdepth.safetensors\",\n    \"hedsketch\": f\"{lora_path}\u002Fhedsketch.safetensors\",\n    \"pose\": f\"{lora_path}\u002Fpose.safetensors\",\n    \"seg\": f\"{lora_path}\u002Fseg.safetensors\",\n    \"inpainting\": f\"{lora_path}\u002Finpainting.safetensors\",\n    \"subject\": f\"{lora_path}\u002Fsubject.safetensors\",\n}\n```\n\n### Single Condition Control\n\n```python\n# Single spatial condition control example\npath = control_models[\"canny\"]\nset_single_lora(pipe.transformer, path, lora_weights=[1], cond_size=512)\n\n# Generate image\nprompt = \"A nice car on the beach\"\nspatial_image = Image.open(\".\u002Fhttps:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_dd888779b6f7.png\").convert(\"RGB\")\n\nimage = pipe(\n    prompt,\n    height=720,\n    width=992,\n    guidance_scale=3.5,\n    num_inference_steps=25,\n    max_sequence_length=512,\n    generator=torch.Generator(\"cpu\").manual_seed(5),\n    spatial_images=[spatial_image],\n    cond_size=512,\n).images[0]\n\n# Clear cache after generation\nclear_cache(pipe.transformer)\n```\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_dd888779b6f7.png\" alt=\"Canny Condition\" width=\"400\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_f89ef8c74b1f.png\" alt=\"Generated Result\" width=\"400\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">Canny Condition\u003C\u002Ftd>\n    \u003Ctd align=\"center\">Generated Result\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n```python\n# Single subject condition control example\npath = control_models[\"subject\"]\nset_single_lora(pipe.transformer, path, lora_weights=[1], cond_size=512)\n\n# Generate image\nprompt = \"A SKS in the library\"\nsubject_image = Image.open(\".\u002Fhttps:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_93f777c8939a.png\").convert(\"RGB\")\n\nimage = pipe(\n    prompt,\n    height=1024,\n    width=1024,\n    guidance_scale=3.5,\n    num_inference_steps=25,\n    max_sequence_length=512,\n    generator=torch.Generator(\"cpu\").manual_seed(5),\n    subject_images=[subject_image],\n    cond_size=512,\n).images[0]\n\n# Clear cache after generation\nclear_cache(pipe.transformer)\n```\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_93f777c8939a.png\" alt=\"Subject Condition\" width=\"400\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_fe600aed5968.png\" alt=\"Generated Result\" width=\"400\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">Subject Condition\u003C\u002Ftd>\n    \u003Ctd align=\"center\">Generated Result\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n### Multi-Condition Control\n\n```python\n# Multi-condition control example\npaths = [control_models[\"subject\"], control_models[\"inpainting\"]]\nset_multi_lora(pipe.transformer, paths, lora_weights=[[1], [1]], cond_size=512)\n\nprompt = \"A SKS on the car\"\nsubject_images = [Image.open(\".\u002Fhttps:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_1ba5048b3d95.png\").convert(\"RGB\")]\nspatial_images = [Image.open(\".\u002Fhttps:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_ca5e99b867e2.png\").convert(\"RGB\")]\n\nimage = pipe(\n    prompt,\n    height=1024,\n    width=1024,\n    guidance_scale=3.5,\n    num_inference_steps=25,\n    max_sequence_length=512,\n    generator=torch.Generator(\"cpu\").manual_seed(42),\n    subject_images=subject_images,\n    spatial_images=spatial_images,\n    cond_size=512,\n).images[0]\n\n# Clear cache after generation\nclear_cache(pipe.transformer)\n```\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_1ba5048b3d95.png\" alt=\"Subject Condition\" width=\"250\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_ca5e99b867e2.png\" alt=\"Inpainting Condition\" width=\"250\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_32302bdcf503.png\" alt=\"Generated Result\" width=\"250\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">Subject Condition\u003C\u002Ftd>\n    \u003Ctd align=\"center\">Inpainting Condition\u003C\u002Ftd>\n    \u003Ctd align=\"center\">Generated Result\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n### Ghibli-Style Portrait Generation\n\n```python\nimport spaces\nimport os\nimport json\nimport time\nimport torch\nfrom PIL import Image\nfrom tqdm import tqdm\nimport gradio as gr\n\nfrom safetensors.torch import save_file\nfrom src.pipeline import FluxPipeline\nfrom src.transformer_flux import FluxTransformer2DModel\nfrom src.lora_helper import set_single_lora, set_multi_lora, unset_lora\n\n# Initialize the image processor\nbase_path = \"black-forest-labs\u002FFLUX.1-dev\"    \nlora_base_path = \".\u002Fcheckpoints\u002Fmodels\"\n\n\npipe = FluxPipeline.from_pretrained(base_path, torch_dtype=torch.bfloat16)\ntransformer = FluxTransformer2DModel.from_pretrained(base_path, subfolder=\"transformer\", torch_dtype=torch.bfloat16)\npipe.transformer = transformer\npipe.to(\"cuda\")\n\ndef clear_cache(transformer):\n    for name, attn_processor in transformer.attn_processors.items():\n        attn_processor.bank_kv.clear()\n\n# Define the Gradio interface\n@spaces.GPU()\ndef single_condition_generate_image(prompt, spatial_img, height, width, seed, control_type):\n    # Set the control type\n    if control_type == \"Ghibli\":\n        lora_path = os.path.join(lora_base_path, \"Ghibli.safetensors\")\n    set_single_lora(pipe.transformer, lora_path, lora_weights=[1], cond_size=512)\n    \n    # Process the image\n    spatial_imgs = [spatial_img] if spatial_img else []\n    image = pipe(\n        prompt,\n        height=int(height),\n        width=int(width),\n        guidance_scale=3.5,\n        num_inference_steps=25,\n        max_sequence_length=512,\n        generator=torch.Generator(\"cpu\").manual_seed(seed), \n        subject_images=[],\n        spatial_images=spatial_imgs,\n        cond_size=512,\n    ).images[0]\n    clear_cache(pipe.transformer)\n    return image\n\n# Define the Gradio interface components\ncontrol_types = [\"Ghibli\"]\n\n\n# Create the Gradio Blocks interface\nwith gr.Blocks() as demo:\n    gr.Markdown(\"# Ghibli Studio Control Image Generation with EasyControl\")\n    gr.Markdown(\"The model is trained on **only 100 real Asian faces** paired with **GPT-4o-generated Ghibli-style counterparts**, and it preserves facial features while applying the iconic anime aesthetic.\")\n    gr.Markdown(\"Generate images using EasyControl with Ghibli control LoRAs.（Due to hardware constraints, only low-resolution images can be generated. For high-resolution (1024+), please set up your own environment.）\")\n    \n    gr.Markdown(\"**[Attention!!]**：The recommended prompts for using Ghibli Control LoRA should include the trigger words: `Ghibli Studio style, Charming hand-drawn anime-style illustration`\")\n    gr.Markdown(\"😊😊If you like this demo, please give us a star (github: [EasyControl](https:\u002F\u002Fgithub.com\u002FXiaojiu-z\u002FEasyControl))\")\n\n    with gr.Tab(\"Ghibli Condition Generation\"):\n        with gr.Row():\n            with gr.Column():\n                prompt = gr.Textbox(label=\"Prompt\", value=\"Ghibli Studio style, Charming hand-drawn anime-style illustration\")\n                spatial_img = gr.Image(label=\"Ghibli Image\", type=\"pil\")  # 上传图像文件\n                height = gr.Slider(minimum=256, maximum=1024, step=64, label=\"Height\", value=768)\n                width = gr.Slider(minimum=256, maximum=1024, step=64, label=\"Width\", value=768)\n                seed = gr.Number(label=\"Seed\", value=42)\n                control_type = gr.Dropdown(choices=control_types, label=\"Control Type\")\n                single_generate_btn = gr.Button(\"Generate Image\")\n            with gr.Column():\n                single_output_image = gr.Image(label=\"Generated Image\")\n\n\n    # Link the buttons to the functions\n    single_generate_btn.click(\n        single_condition_generate_image,\n        inputs=[prompt, spatial_img, height, width, seed, control_type],\n        outputs=single_output_image\n    )\n\n# Launch the Gradio app\ndemo.queue().launch()\n```\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_9daa8f19d1b7.png\" alt=\"Input Image\" width=\"250\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_946af6c4a6ee.png\" alt=\"Generated Result\" width=\"250\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">Input Image\u003C\u002Ftd>\n    \u003Ctd align=\"center\">Generated Result\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n## Usage Tips\n\n- Clear cache after each generation using `clear_cache(pipe.transformer)`\n- For optimal performance:\n  - Start with `guidance_scale=3.5` and adjust based on results\n  - Use `num_inference_steps=25` for a good balance of quality and speed\n- When using set_multi_lora api, make sure the subject lora path(subject) is before the spatial lora path(canny, depth, hedsketch, etc.).\n\n## Todo List\n1. - [x] Inference code \n2. - [x] Spatial Pre-trained weights \n3. - [x] Subject Pre-trained weights \n4. - [x] Training code\n\n\n## Star History\n\n[![Star History Chart](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_87402e375c0c.png)](https:\u002F\u002Fstar-history.com\u002F#Xiaojiu-z\u002FEasyControl&Date)\n\n## Disclaimer\nThe code of EasyControl is released under [Apache License](https:\u002F\u002Fgithub.com\u002FXiaojiu-Z\u002FEasyControl?tab=Apache-2.0-1-ov-file#readme) for both academic and commercial usage. Our released checkpoints are for research purposes only. Users are granted the freedom to create images using this tool, but they are obligated to comply with local laws and utilize it responsibly. The developers will not assume any responsibility for potential misuse by users.\n\n## Hiring\u002FCooperation\n- **2025-04-03**: If you are interested in EasyControl and beyond, or if you are interested in building 4o-like capacity (in a low-budget way, of course), we can collaborate offline in Shanghai\u002FBeijing\u002FHong Kong\u002FSingapore or online.\nReach out: **jmliu1217@gmail.com (wechat: jiaming068870)**\n\n## Citation\n```bibtex\n@article{zhang2025easycontrol,\n  title={EasyControl: Adding Efficient and Flexible Control for Diffusion Transformer},\n  author={Zhang, Yuxuan and Yuan, Yirui and Song, Yiren and Wang, Haofan and Liu, Jiaming},\n  journal={arXiv preprint arXiv:2503.07027},\n  year={2025}\n}\n```\n","# EasyControl 的实现\n\nEasyControl：为扩散 Transformer 添加高效灵活的控制机制\n\n\u003Ca href='https:\u002F\u002Feasycontrolproj.github.io\u002F'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject-Page-green'>\u003C\u002Fa>\n\u003Ca href='https:\u002F\u002Farxiv.org\u002Fpdf\u002F2503.07027'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FTechnique-Report-red'>\u003C\u002Fa> \n\u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002FXiaojiu-Z\u002FEasyControl\u002F\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F🤗_HuggingFace-Model-ffbd45.svg\" alt=\"HuggingFace\">\u003C\u002Fa>\n\u003Ca href='https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fjamesliu1217\u002FEasyControl'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20Hugging%20Face-Spaces-blue'>\u003C\u002Fa>\n\u003Ca href='https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fjamesliu1217\u002FEasyControl_Ghibli'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20Hugging%20Face-Ghibli_Spaces-purple'>\u003C\u002Fa>\n\n> *[张宇轩](https:\u002F\u002Fxiaojiu-z.github.io\u002FYuxuanZhang.github.io\u002F)、[袁一睿](https:\u002F\u002Fgithub.com\u002FReynoldyy)、[宋怡仁](https:\u002F\u002Fscholar.google.com.hk\u002Fcitations?user=L2YS0jgAAAAJ)、[王浩凡](https:\u002F\u002Fhaofanwang.github.io\u002F)、[刘嘉铭](https:\u002F\u002Fscholar.google.com\u002Fcitations?user=SmL7oMQAAAAJ&hl=en)]*\n> \u003Cbr>\n> 香港中文大学、上海科技大学、Tiamat AI、阿里巴巴集团、新加坡国立大学、Liblib AI\n\n\u003Cimg src='https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_3a80e064e594.jpg'>\n\n## 特性\n* **动机:** 扩散模型的架构正从基于 Unet 转向 DiT（扩散 Transformer）。然而，DiT 生态系统缺乏成熟的插件支持，并面临效率瓶颈、多条件协调冲突以及模型适应性不足等挑战。\n* **贡献:** 我们提出了 EasyControl，一个高效且灵活的统一条件 DiT 框架。通过引入轻量级条件注入 LoRA 模块、位置感知训练范式，以及因果注意力机制与 KV 缓存技术的结合，我们显著提升了 **模型兼容性**（实现即插即用功能和风格无损控制）、**生成灵活性**（支持多种分辨率、宽高比及多条件组合）和 **推理效率**。\n\u003Cimg src='https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_c6d83e10bf5f.jpg'>\n\n## 新闻\n- **2025-04-11**: 🔥🔥🔥 训练代码已发布。推荐硬件：至少 1 张 NVIDIA H100\u002FH800\u002FA100 显卡，显存约 80GB。\n- **2025-04-09**: ⭐️ 简易 API 的代码已发布。如果您希望在个人设备上运行这些模型，请前往 simple_api 分支获取相关资源。\n\n- **2025-04-07**: 🔥 感谢 [CFG-Zero*](https:\u002F\u002Fgithub.com\u002FWeichenFan\u002FCFG-Zero-star) 团队的出色工作，EasyControl 现已与 CFG-Zero* 集成！！只需几行代码，您就可以大幅提升图像质量和可控性！！您可以从 [此链接](https:\u002F\u002Fgithub.com\u002FWeichenFan\u002FCFG-Zero-star\u002Fblob\u002Fmain\u002Fmodels\u002Feasycontrol\u002Finfer.py) 下载修改后的代码并尝试。\n\n\u003Ctable class=\"center\">\n  \u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_d05c5684cb60.webp\" style=\"width:410px; height:auto;\">\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_22c4d532bd34.webp\" style=\"width:410px; height:auto;\">\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_2f830fe19776.webp\" style=\"width:410px; height:auto;\">\u003C\u002Ftd>\n  \u003C\u002Ftr>\n  \u003Ctr>\n    \u003Ctd align=\"center\">\u003Cb>源图像\u003C\u002Fb>\u003C\u002Ftd>\n    \u003Ctd align=\"center\">\u003Cb>CFG\u003C\u002Fb>\u003C\u002Ftd>\n    \u003Ctd align=\"center\">\u003Cb>CFG-Zero*\u003C\u002Fb>\u003C\u002Ftd>\n  \u003C\u002Ftr>\n\u003C\u002Ftable>\n\n- **2025-04-03**: 感谢 [jax-explorer](https:\u002F\u002Fgithub.com\u002Fjax-explorer)，现已支持 [Ghibli Img2Img 控制 ComfyUI 节点](https:\u002F\u002Fgithub.com\u002Fjax-explorer\u002FComfyUI-easycontrol)！\n\n- **2025-04-01**: 🔥 新的 [风格化 Img2Img 控制模型](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fjamesliu1217\u002FEasyControl_Ghibli)现已发布！！使用此 LoRA 模型将人像转化为吉卜力工作室风格的艺术作品。该模型仅基于 **100 张真实的亚洲人脸** 和 **由 GPT-4o 生成的吉卜力风格对应图** 进行训练，能够在保留面部特征的同时应用标志性的动漫美学。\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_94779297bdc3.jpeg\" alt=\"示例 3\" width=\"400\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_0982b737852d.jpeg\" alt=\"示例 4\" width=\"400\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">示例 3\u003C\u002Ftd>\n    \u003Ctd align=\"center\">示例 4\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n- **2025-03-19**: 🔥 我们已在 Hugging Face 上发布了 [演示空间](https:\u002F\u002Fhuggingface.co\u002Fspaces\u002Fjamesliu1217\u002FEasyControl)! 您现在可以通过 Hugging Face 空间试用 EasyControl，尽情享受吧！\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_8bc7770e095f.jpeg\" alt=\"示例 1\" width=\"400\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_04e8860a7b68.jpeg\" alt=\"示例 2\" width=\"400\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">示例 1\u003C\u002Ftd>\n    \u003Ctd align=\"center\">示例 2\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n- **2025-03-18**: 🔥 我们已在 Hugging Face 上发布了 [预训练检查点](https:\u002F\u002Fhuggingface.co\u002FXiaojiu-Z\u002FEasyControl\u002F)！您现在可以使用官方权重试用 EasyControl。\n- **2025-03-12**: ⭐️ 推理代码已发布。待确认一切正常后，新模型将合并到本仓库中。敬请期待后续更新！😊\n\n## 安装\n\n我们建议使用 Python 3.10 和支持 CUDA 的 PyTorch。搭建环境的步骤如下：\n\n```bash\n# 创建一个新的 conda 环境\nconda create -n easycontrol python=3.10\nconda activate easycontrol\n\n# 安装其他依赖\npip install -r requirements.txt\n```\n\n## 下载\n\n您可以直接从 [Hugging Face](https:\u002F\u002Fhuggingface.co\u002FEasyControl\u002FEasyControl) 下载模型。\n或者使用 Python 脚本下载：\n\n```python\nfrom huggingface_hub import hf_hub_download\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fcanny.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fdepth.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fhedsketch.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Finpainting.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fpose.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fseg.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002Fsubject.safetensors\", local_dir=\".\u002F\")\nhf_hub_download(repo_id=\"Xiaojiu-Z\u002FEasyControl\", filename=\"models\u002FGhibli.safetensors\", local_dir=\".\u002F\")\n```\n\n如果您无法访问 Hugging Face，可以使用 [hf-mirror](https:\u002F\u002Fhf-mirror.com\u002F) 下载模型：\n```python\nexport HF_ENDPOINT=https:\u002F\u002Fhf-mirror.com\nhuggingface-cli download --resume-download Xiaojiu-Z\u002FEasyControl --local-dir checkpoints --local-dir-use-symlinks False\n```\n\n## 使用方法\n以下是使用 EasyControl 的基本示例：\n\n### 模型初始化\n\n```python\nimport torch\nfrom PIL import Image\nfrom src.pipeline import FluxPipeline\nfrom src.transformer_flux import FluxTransformer2DModel\nfrom src.lora_helper import set_single_lora, set_multi_lora\n\ndef clear_cache(transformer):\n    for name, attn_processor in transformer.attn_processors.items():\n        attn_processor.bank_kv.clear()\n\n# 初始化模型\ndevice = \"cuda\"\nbase_path = \"FLUX.1-dev\"  # 您的基础模型路径\npipe = FluxPipeline.from_pretrained(base_path, torch_dtype=torch.bfloat16, device=device)\ntransformer = FluxTransformer2DModel.from_pretrained(\n    base_path, \n    subfolder=\"transformer\",\n    torch_dtype=torch.bfloat16, \n    device=device\n)\npipe.transformer = transformer\npipe.to(device)\n\n# 加载控制模型\nlora_path = \".\u002Fcheckpoints\u002Fmodels\"\ncontrol_models = {\n    \"canny\": f\"{lora_path}\u002Fcanny.safetensors\",\n    \"depth\": f\"{lora_path}\u002Fdepth.safetensors\",\n    \"hedsketch\": f\"{lora_path}\u002Fhedsketch.safetensors\",\n    \"pose\": f\"{lora_path}\u002Fpose.safetensors\",\n    \"seg\": f\"{lora_path}\u002Fseg.safetensors\",\n    \"inpainting\": f\"{lora_path}\u002Finpainting.safetensors\",\n    \"subject\": f\"{lora_path}\u002Fsubject.safetensors\",\n}\n```\n\n### 单条件控制\n\n```python\n# 单空间条件控制示例\npath = control_models[\"canny\"]\nset_single_lora(pipe.transformer, path, lora_weights=[1], cond_size=512)\n\n# 生成图像\nprompt = \"海滩上的一辆漂亮的汽车\"\nspatial_image = Image.open(\".\u002Fhttps:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_dd888779b6f7.png\").convert(\"RGB\")\n\nimage = pipe(\n    prompt,\n    height=720,\n    width=992,\n    guidance_scale=3.5,\n    num_inference_steps=25,\n    max_sequence_length=512,\n    generator=torch.Generator(\"cpu\").manual_seed(5),\n    spatial_images=[spatial_image],\n    cond_size=512,\n).images[0]\n\n# 生成后清除缓存\nclear_cache(pipe.transformer)\n```\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_dd888779b6f7.png\" alt=\"Canny 条件\" width=\"400\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_f89ef8c74b1f.png\" alt=\"生成结果\" width=\"400\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">Canny 条件\u003C\u002Ftd>\n    \u003Ctd align=\"center\">生成结果\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n```python\n# 单主体条件控制示例\npath = control_models[\"subject\"]\nset_single_lora(pipe.transformer, path, lora_weights=[1], cond_size=512)\n\n# 生成图像\nprompt = \"图书馆里的SKS步枪\"\nsubject_image = Image.open(\".\u002Fhttps:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_93f777c8939a.png\").convert(\"RGB\")\n\nimage = pipe(\n    prompt,\n    height=1024,\n    width=1024,\n    guidance_scale=3.5,\n    num_inference_steps=25,\n    max_sequence_length=512,\n    generator=torch.Generator(\"cpu\").manual_seed(5),\n    subject_images=[subject_image],\n    cond_size=512,\n).images[0]\n\n# 生成后清除缓存\nclear_cache(pipe.transformer)\n```\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_93f777c8939a.png\" alt=\"主体条件\" width=\"400\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_fe600aed5968.png\" alt=\"生成结果\" width=\"400\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">主体条件\u003C\u002Ftd>\n    \u003Ctd align=\"center\">生成结果\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n### 多条件控制\n\n```python\n# 多条件控制示例\npaths = [control_models[\"subject\"], control_models[\"inpainting\"]]\nset_multi_lora(pipe.transformer, paths, lora_weights=[[1], [1]], cond_size=512)\n\nprompt = \"车上的SKS步枪\"\nsubject_images = [Image.open(\".\u002Fhttps:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_1ba5048b3d95.png\").convert(\"RGB\")]\nspatial_images = [Image.open(\".\u002Fhttps:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_ca5e99b867e2.png\").convert(\"RGB\")]\n\nimage = pipe(\n    prompt,\n    height=1024,\n    width=1024,\n    guidance_scale=3.5,\n    num_inference_steps=25,\n    max_sequence_length=512,\n    generator=torch.Generator(\"cpu\").manual_seed(42),\n    subject_images=subject_images,\n    spatial_images=spatial_images,\n    cond_size=512,\n).images[0]\n\n# 生成后清除缓存\nclear_cache(pipe.transformer)\n```\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_1ba5048b3d95.png\" alt=\"主体条件\" width=\"250\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_ca5e99b867e2.png\" alt=\"修复条件\" width=\"250\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_32302bdcf503.png\" alt=\"生成结果\" width=\"250\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">主体条件\u003C\u002Ftd>\n    \u003Ctd align=\"center\">修复条件\u003C\u002Ftd>\n    \u003Ctd align=\"center\">生成结果\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n### 吉卜力风格人像生成\n\n```python\nimport spaces\nimport os\nimport json\nimport time\nimport torch\nfrom PIL import Image\nfrom tqdm import tqdm\nimport gradio as gr\n\nfrom safetensors.torch import save_file\nfrom src.pipeline import FluxPipeline\nfrom src.transformer_flux import FluxTransformer2DModel\nfrom src.lora_helper import set_single_lora, set_multi_lora, unset_lora\n\n# 初始化图像处理器\nbase_path = \"black-forest-labs\u002FFLUX.1-dev\"    \nlora_base_path = \".\u002Fcheckpoints\u002Fmodels\"\n\n\npipe = FluxPipeline.from_pretrained(base_path, torch_dtype=torch.bfloat16)\ntransformer = FluxTransformer2DModel.from_pretrained(base_path, subfolder=\"transformer\", torch_dtype=torch.bfloat16)\npipe.transformer = transformer\npipe.to(\"cuda\")\n\ndef clear_cache(transformer):\n    for name, attn_processor in transformer.attn_processors.items():\n        attn_processor.bank_kv.clear()\n\n# 定义 Gradio 界面\n@spaces.GPU()\ndef single_condition_generate_image(prompt, spatial_img, height, width, seed, control_type):\n    # 设置控制类型\n    if control_type == \"Ghibli\":\n        lora_path = os.path.join(lora_base_path, \"Ghibli.safetensors\")\n    set_single_lora(pipe.transformer, lora_path, lora_weights=[1], cond_size=512)\n    \n    # 处理图像\n    spatial_imgs = [spatial_img] if spatial_img else []\n    image = pipe(\n        prompt,\n        height=int(height),\n        width=int(width),\n        guidance_scale=3.5,\n        num_inference_steps=25,\n        max_sequence_length=512,\n        generator=torch.Generator(\"cpu\").manual_seed(seed), \n        subject_images=[],\n        spatial_images=spatial_imgs,\n        cond_size=512,\n    ).images[0]\n    clear_cache(pipe.transformer)\n    return image\n\n# 定义 Gradio 界面组件\ncontrol_types = [\"Ghibli\"]\n\n# 创建 Gradio Blocks 界面\nwith gr.Blocks() as demo:\n    gr.Markdown(\"# 使用 EasyControl 控制吉卜力工作室风格图像生成\")\n    gr.Markdown(\"该模型仅基于**100张真实的亚洲人脸**与**由 GPT-4o 生成的吉卜力风格对应图像**进行训练，在保留面部特征的同时，能够应用标志性的动漫美学风格。\")\n    gr.Markdown(\"使用带有吉卜力控制 LoRA 的 EasyControl 生成图像。（由于硬件限制，目前只能生成低分辨率图像。如需高分辨率（1024+），请自行搭建环境。）\")\n    \n    gr.Markdown(\"**[注意！！]**：推荐用于吉卜力控制 LoRA 的提示词应包含触发词：`吉卜力工作室风格、迷人手绘动漫风格插画`\")\n    gr.Markdown(\"😊😊如果您喜欢这个演示，请给我们点个赞（GitHub：[EasyControl](https:\u002F\u002Fgithub.com\u002FXiaojiu-z\u002FEasyControl))\")\n\n    with gr.Tab(\"吉卜力条件生成\"):\n        with gr.Row():\n            with gr.Column():\n                prompt = gr.Textbox(label=\"提示词\", value=\"吉卜力工作室风格、迷人手绘动漫风格插画\")\n                spatial_img = gr.Image(label=\"吉卜力图像\", type=\"pil\")  # 上传图像文件\n                height = gr.Slider(minimum=256, maximum=1024, step=64, label=\"高度\", value=768)\n                width = gr.Slider(minimum=256, maximum=1024, step=64, label=\"宽度\", value=768)\n                seed = gr.Number(label=\"种子\", value=42)\n                control_type = gr.Dropdown(choices=control_types, label=\"控制类型\")\n                single_generate_btn = gr.Button(\"生成图像\")\n            with gr.Column():\n                single_output_image = gr.Image(label=\"生成图像\")\n\n\n    # 将按钮与函数关联\n    single_generate_btn.click(\n        single_condition_generate_image,\n        inputs=[prompt, spatial_img, height, width, seed, control_type],\n        outputs=single_output_image\n    )\n\n# 启动 Gradio 应用\ndemo.queue().launch()\n```\n\n\u003Cdiv align=\"center\">\n\u003Ctable>\n\u003Ctr>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_9daa8f19d1b7.png\" alt=\"输入图像\" width=\"250\"\u002F>\u003C\u002Ftd>\n    \u003Ctd>\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_946af6c4a6ee.png\" alt=\"生成结果\" width=\"250\"\u002F>\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n    \u003Ctd align=\"center\">输入图像\u003C\u002Ftd>\n    \u003Ctd align=\"center\">生成结果\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\n## 使用提示\n\n- 每次生成后请使用 `clear_cache(pipe.transformer)` 清除缓存。\n- 为获得最佳效果：\n  - 建议从 `guidance_scale=3.5` 开始，并根据结果调整。\n  - 使用 `num_inference_steps=25` 可在质量和速度之间取得良好平衡。\n- 使用 `set_multi_lora` API 时，请确保主体 LoRA 路径（subject）位于空间 LoRA 路径（canny、depth、hedsketch 等）之前。\n\n## 待办事项清单\n1. - [x] 推理代码 \n2. - [x] 空间预训练权重 \n3. - [x] 主体预训练权重 \n4. - [x] 训练代码\n\n\n## 星标历史\n\n[![星标历史图表](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_readme_87402e375c0c.png)](https:\u002F\u002Fstar-history.com\u002F#Xiaojiu-z\u002FEasyControl&Date)\n\n## 免责声明\nEasyControl 的代码采用 [Apache 许可证](https:\u002F\u002Fgithub.com\u002FXiaojiu-Z\u002FEasyControl?tab=Apache-2.0-1-ov-file#readme) 发布，适用于学术和商业用途。我们发布的检查点仅供研究目的使用。用户可以自由地使用此工具生成图像，但必须遵守当地法律并负责任地使用。开发者对用户的任何潜在滥用行为不承担任何责任。\n\n## 招聘\u002F合作\n- **2025年4月3日**：如果您对 EasyControl 及其相关技术感兴趣，或者希望以低成本方式构建类似 4o 的能力，我们可以在上海、北京、香港、新加坡等地线下合作，或通过线上方式进行交流。\n联系方式：**jmliu1217@gmail.com（微信：jiaming068870）**\n\n## 引用\n```bibtex\n@article{zhang2025easycontrol,\n  title={EasyControl: 为扩散 Transformer 添加高效灵活的控制},\n  author={Zhang, Yuxuan 和 Yuan, Yirui 和 Song, Yiren 和 Wang, Haofan 和 Liu, Jiaming},\n  journal={arXiv 预印本 arXiv:2503.07027},\n  year={2025}\n}\n```","# EasyControl 快速上手指南\n\nEasyControl 是一个高效且灵活的扩散 Transformer (DiT) 控制框架，旨在为 FLUX.1 等模型提供插件式的条件控制（如边缘、深度、姿态、主体保持等），支持多条件组合与高分辨率生成。\n\n## 环境准备\n\n*   **操作系统**: Linux \u002F Windows (推荐 Linux)\n*   **Python 版本**: 3.10\n*   **GPU 要求**: 支持 CUDA 的 NVIDIA 显卡\n    *   **推理**: 建议显存 ≥ 24GB (取决于分辨率)\n    *   **训练**: 建议至少 1x NVIDIA H100\u002FH800\u002FA100 (约 80GB 显存)\n*   **前置依赖**: PyTorch (带 CUDA 支持), Conda (推荐)\n\n## 安装步骤\n\n### 1. 创建虚拟环境\n推荐使用 Conda 创建独立的 Python 3.10 环境：\n\n```bash\nconda create -n easycontrol python=3.10\nconda activate easycontrol\n```\n\n### 2. 安装依赖库\n进入项目目录后安装所需包：\n\n```bash\npip install -r requirements.txt\n```\n\n### 3. 下载模型权重\n你可以从 Hugging Face 下载预训练模型。如果直接访问受限，**强烈推荐使用国内镜像源加速下载**。\n\n**方式一：使用 Python 脚本下载 (推荐)**\n```python\nfrom huggingface_hub import hf_hub_download\n\n# 定义需要下载的模型列表\nmodels = [\n    \"canny.safetensors\", \"depth.safetensors\", \"hedsketch.safetensors\",\n    \"inpainting.safetensors\", \"pose.safetensors\", \"seg.safetensors\",\n    \"subject.safetensors\", \"Ghibli.safetensors\"\n]\n\nfor model_name in models:\n    hf_hub_download(\n        repo_id=\"Xiaojiu-Z\u002FEasyControl\", \n        filename=f\"models\u002F{model_name}\", \n        local_dir=\".\u002Fcheckpoints\"\n    )\n```\n\n**方式二：使用命令行 + 国内镜像 (hf-mirror)**\n如果无法访问官方源，请设置镜像地址后下载：\n\n```bash\nexport HF_ENDPOINT=https:\u002F\u002Fhf-mirror.com\nhuggingface-cli download --resume-download Xiaojiu-Z\u002FEasyControl --local-dir checkpoints --local-dir-use-symlinks False\n```\n\n> **注意**: 使用前请确保已下载基础模型 `FLUX.1-dev`，并在代码中指定其路径。\n\n## 基本使用\n\n以下是最简单的单条件控制（以 Canny 边缘检测为例）生成流程。\n\n### 1. 初始化模型\n加载 FLUX.1 基础模型并注入 EasyControl 组件。\n\n```python\nimport torch\nfrom PIL import Image\nfrom src.pipeline import FluxPipeline\nfrom src.transformer_flux import FluxTransformer2DModel\nfrom src.lora_helper import set_single_lora\n\ndef clear_cache(transformer):\n    for name, attn_processor in transformer.attn_processors.items():\n        attn_processor.bank_kv.clear()\n\n# 配置设备与路径\ndevice = \"cuda\"\nbase_path = \"FLUX.1-dev\"  # 替换为你本地的 FLUX.1-dev 模型路径\nlora_path = \".\u002Fcheckpoints\u002Fmodels\"\n\n# 初始化 Pipeline 和 Transformer\npipe = FluxPipeline.from_pretrained(base_path, torch_dtype=torch.bfloat16, device=device)\ntransformer = FluxTransformer2DModel.from_pretrained(\n    base_path, \n    subfolder=\"transformer\",\n    torch_dtype=torch.bfloat16, \n    device=device\n)\npipe.transformer = transformer\npipe.to(device)\n\n# 加载控制模型 (例如 Canny)\ncontrol_model_path = f\"{lora_path}\u002Fcanny.safetensors\"\nset_single_lora(pipe.transformer, control_model_path, lora_weights=[1], cond_size=512)\n```\n\n### 2. 生成图像\n准备一张条件图（如边缘图），输入提示词进行生成。\n\n```python\n# 准备输入\nprompt = \"A nice car on the beach\"\nspatial_image = Image.open(\".\u002Ftest_imgs\u002Fcanny.png\").convert(\"RGB\") # 替换为你的条件图路径\n\n# 执行生成\nimage = pipe(\n    prompt,\n    height=720,\n    width=992,\n    guidance_scale=3.5,\n    num_inference_steps=25,\n    max_sequence_length=512,\n    generator=torch.Generator(\"cpu\").manual_seed(5),\n    spatial_images=[spatial_image], # 传入条件图像列表\n    cond_size=512,\n).images[0]\n\n# 生成完成后清理缓存 (重要)\nclear_cache(pipe.transformer)\n\n# 保存结果\nimage.save(\"result.png\")\n```\n\n### 3. 多条件控制 (可选)\nEasyControl 支持同时叠加多个条件（例如：主体保持 + 局部重绘）。\n\n```python\nfrom src.lora_helper import set_multi_lora\n\n# 加载多个 LoRA (例如：主体 + 重绘)\npaths = [f\"{lora_path}\u002Fsubject.safetensors\", f\"{lora_path}\u002Finpainting.safetensors\"]\nset_multi_lora(pipe.transformer, paths, lora_weights=[[1], [1]], cond_size=512)\n\nprompt = \"A SKS on the car\"\nsubject_images = [Image.open(\".\u002Ftest_imgs\u002Fsubject_1.png\").convert(\"RGB\")]\nspatial_images = [Image.open(\".\u002Ftest_imgs\u002Finpainting.png\").convert(\"RGB\")]\n\nimage = pipe(\n    prompt,\n    height=1024,\n    width=1024,\n    guidance_scale=3.5,\n    num_inference_steps=25,\n    max_sequence_length=512,\n    generator=torch.Generator(\"cpu\").manual_seed(42),\n    subject_images=subject_images,\n    spatial_images=spatial_images,\n    cond_size=512,\n).images[0]\n\nclear_cache(pipe.transformer)\nimage.save(\"multi_condition_result.png\")\n```","一家专注于二次元风格化的游戏美术团队，正试图将大量真人角色肖像快速转化为吉卜力风格的设定图，同时严格保留人物面部特征。\n\n### 没有 EasyControl 时\n- **风格与特征难以兼得**：传统图生图方法在施加强烈吉卜力滤镜时，极易丢失原图人物的五官细节，导致“换脸”失败。\n- **多条件控制冲突**：当同时输入姿态骨架、边缘检测和文本提示时，基于 Unet 的旧模型常出现条件互相干扰，生成画面崩坏。\n- **推理成本高昂**：处理高分辨率或多长宽比图片时，显存占用激增，单张渲染耗时过长，严重拖慢迭代效率。\n- **模型适配困难**：缺乏针对 DiT 架构的成熟插件，每次尝试新风格都需要重新训练庞大模型，无法实现“即插即用”。\n\n### 使用 EasyControl 后\n- **无损风格迁移**：利用轻量级 Condition Injection LoRA 模块，仅用少量数据即可在完美保留人脸特征的前提下，精准施加吉卜力画风。\n- **灵活的多条件协同**：凭借位置感知训练范式，轻松融合姿态、边缘等多重控制信号，画面结构稳定且逻辑自洽。\n- **高效推理体验**：结合因果注意力机制与 KV Cache 技术，显著降低显存需求并提升生成速度，支持任意分辨率和长宽比输出。\n- **插件化工作流**：直接通过 ComfyUI 节点调用预训练模型，无需重新训练即可切换不同风格控制，真正实现模块化生产。\n\nEasyControl 通过统一的 DiT 控制框架，解决了高保真风格迁移中的效率与兼容性难题，让创意落地不再受限于技术瓶颈。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FXiaojiu-z_EasyControl_3a80e064.jpg","Xiaojiu-z",null,"https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002FXiaojiu-z_615305e6.jpg","CUHK","Hong Kong","https:\u002F\u002Fgithub.com\u002FXiaojiu-z",[82,86,90],{"name":83,"color":84,"percentage":85},"Python","#3572A5",96.6,{"name":87,"color":88,"percentage":89},"Jupyter Notebook","#DA5B0B",2.1,{"name":91,"color":92,"percentage":93},"Shell","#89e051",1.3,1725,126,"2026-04-03T17:24:07","Apache-2.0","Linux","训练必需：至少 1x NVIDIA H100\u002FH800\u002FA100，显存约 80GB；推理必需：支持 CUDA 的 NVIDIA GPU（代码示例指定 device='cuda'），具体显存未说明但需加载 FLUX.1-dev 基座模型及多个 LoRA，建议大显存","未说明",{"notes":102,"python":103,"dependencies":104},"1. 该工具基于 Diffusion Transformer (DiT) 架构，目前主要明确支持 Linux 环境（通过 conda 安装）。2. 训练代码对硬件要求极高，官方推荐至少一张 80GB 显存的 H100\u002FA100 显卡。3. 推理依赖 FLUX.1-dev 作为基座模型，需单独下载或配置路径。4. 提供了多种控制模式（边缘、深度、姿态、重绘、主体保持等）及吉卜力风格微调模型。5. 若无法访问 Hugging Face，可使用 hf-mirror 镜像下载模型文件。6. 生成后需手动调用 clear_cache 清理 KV Cache 以避免显存泄漏。","3.10",[105,106,107,108,109,110,111],"torch (CUDA support)","PIL (Pillow)","huggingface_hub","safetensors","gradio","tqdm","spaces",[14,13],"2026-03-27T02:49:30.150509","2026-04-06T08:45:23.779668",[116,121,126,131,136,141,146],{"id":117,"question_zh":118,"answer_zh":119,"source_url":120},9995,"运行 EasyControl 需要多少显存（VRAM）？24GB 显卡能跑吗？","原生 Flux.1 Dev 模型占用少于 24GB 显存，但结合 EasyControl 和一个 LoRA 后，显存需求会增至约 35GB（单个 LoRA 占用超过 10GB）。如果使用高分辨率图片作为输入源，生成单张图片可能需要数小时。对于 24GB 显存的显卡（如 3090\u002F4090），直接运行可能会非常慢或显存不足，建议尝试量化模型或降低分辨率。","https:\u002F\u002Fgithub.com\u002FXiaojiu-z\u002FEasyControl\u002Fissues\u002F22",{"id":122,"question_zh":123,"answer_zh":124,"source_url":125},9996,"如何使用普通的风格 LoRA（如 FLUX-dev-lora-children-simple-sketch）进行风格迁移？","项目提供的 Ghibli LoRA 是专门训练的 EasyControl 模型，与普通用于“文生图”的风格 LoRA 不同。如果你想使用其他普通风格 LoRA（如 children-simple-sketch）来实现图片风格迁移并保持内容一致，不能直接加载使用，你需要基于该风格数据集重新训练一个新的 EasyControl LoRA 模型。","https:\u002F\u002Fgithub.com\u002FXiaojiu-z\u002FEasyControl\u002Fissues\u002F41",{"id":127,"question_zh":128,"answer_zh":129,"source_url":130},9997,"推理时出现黑屏或输出为 NaN 错误怎么办？","这通常是由于 PyTorch 版本兼容性导致的。如果你使用的是 torch 2.5.0 并遇到 bfloat16 或 float16 精度下的黑屏\u002FNaN 问题，请将 PyTorch 升级到 2.6 版本即可解决。如果无法升级，临时解决方案是使用完整的 FP32 精度运行，但这会显著增加推理时间。","https:\u002F\u002Fgithub.com\u002FXiaojiu-z\u002FEasyControl\u002Fissues\u002F32",{"id":132,"question_zh":133,"answer_zh":134,"source_url":135},9998,"训练时遇到显存溢出（OOM）错误如何解决？","如果在高性能显卡（如 H100 80GB）上训练仍遇到 OOM，可以尝试降低条件图像的尺寸参数。建议在训练配置中将 `noise_size` 设置为 768 或更低来进行调试，以减少显存占用。","https:\u002F\u002Fgithub.com\u002FXiaojiu-z\u002FEasyControl\u002Fissues\u002F57",{"id":137,"question_zh":138,"answer_zh":139,"source_url":140},9999,"为什么官方提供的 Ghibli LoRA 在图像反演（Image Inversion）工作流中效果不佳？","官方模型是针对 EasyControl 流程优化的。如果在其他工作流（如普通 ComfyUI 节点）中效果不好，建议尝试使用专为 EasyControl 设计的 ComfyUI 自定义节点（ComfyUI-easycontrol），以确保模型能正确调用其特定的控制机制，从而获得最佳效果。","https:\u002F\u002Fgithub.com\u002FXiaojiu-z\u002FEasyControl\u002Fissues\u002F18",{"id":142,"question_zh":143,"answer_zh":144,"source_url":145},10000,"如何训练人脸一致性（Face Consistent）LoRA？有公开数据集吗？","论文中展示的人脸一致性效果是基于私有数据集训练的，由于版权和安全限制，该数据集未开源。Subject LoRA 与人脸一致性 LoRA 是不同的模型。如果你想自己训练人脸一致性 LoRA，需要收集大规模的多视角人脸数据集，并参考项目中的 `train_subject.sh` 脚本进行类似训练。目前的公开 Subject LoRA 尚无法完美生成一致性人脸。","https:\u002F\u002Fgithub.com\u002FXiaojiu-z\u002FEasyControl\u002Fissues\u002F53",{"id":147,"question_zh":148,"answer_zh":149,"source_url":150},10001,"项目的训练代码已经开源了吗？","是的，训练代码已经发布。用户可以直接在仓库中查找并尝试运行训练脚本。","https:\u002F\u002Fgithub.com\u002FXiaojiu-z\u002FEasyControl\u002Fissues\u002F8",[]]