一、实验引言1.实验目的熟悉千里云平台中ComfyUI节点式工作流的整体结构与核心节点作用掌握基础文生图工作流搭建、参数设置、正负提示词撰写以及图像生成完整流程。理解LoRA模型加载节点的功能学会单LoRA与多LoRA组合使用方式对比基础模型和LoRA增强模型的出图效果。同时探究采样器、采样步数、图像分辨率等参数对图像质量与程序运行效率造成的影响。二、实验环境与软硬件配置千里云容器实例ComfyUI在线运行环境访问地址https://4d8fef1c-8188fz.suanlihou.com三、实验原理ComfyUI采用节点串联的形式完成AI图像生成底层依托Stable Diffusion模型运行。基础大模型决定图像整体画风与画质基底空Latent图像会生成带有噪声的潜在图像空间也是最终画面的原始载体分辨率参数直接决定输出图片尺寸。CLIP文本编码能够将人为编写的正向、负向提示词转化为模型可识别的特征向量正向词用来定义画面内容、风格与细节负向词用于规避畸形、模糊、水印等缺陷。K采样器是核心推理模块依靠采样算法逐步去除噪声采样步数、CFG数值、采样器类型会同时影响画面细节和生成速度。VAE解码负责将处理后的潜在图像转化为常规像素图像最后由保存图像节点导出成品。LoRA属于轻量化微调模型依附基础大模型使用可强化特定角色、画风与装饰细节多个LoRA叠加还能实现多种风格融合。四、实验全流程步骤本次实验分为基础文生图、LoRA模型增强图像生成两大模块按照流程分步开展操作。一前期准备登录千里云容器实例通过指定网址进入ComfyUI操作界面。清空原有工作流依次添加并连接标准文生图所需节点顺序为Checkpoint加载器、空Latent图像、正负向CLIP文本编码、K采样器、VAE解码、保存图像。在模型加载节点中选定指定基础模型统一初始图像分辨率为512×512。二基础文生图实验第一组林间萤火木屋治愈森系风景第二组古风执扇江南少女国风人物第三组星际太空花园科幻幻想风第四组复古街角咖啡店欧式日常场景三LoRA模型增强实验在基础工作流的基础上新增LoRA加载器节点将其串联在基础模型加载节点五、实验结果与详细分析正向提示词 plaintext A small wooden house deep in the forest, twilight, countless fireflies flying around, dense green plants, old wooden windows, warm candlelight through the window, soft mist, peaceful atmosphere, natural scenery, soft focus, masterpiece, best quality, highly detailed 负向提示词通用防缺陷 plaintext lowres, bad anatomy, blurry, text, watermark, logo, signature, ugly, deformed, disfigured, extra objects, messy composition, worst quality, low quality, 3d render, cartoon distortion正向提示词 plaintext 1girl, traditional Chinese style, ancient costume, long hanfu, holding silk fan, long black hair tied with hairpin, standing by lotus pond, willow trees, rippling water, soft breeze, elegant posture, classical beauty, ink painting style, delicate facial features, vibrant colors, masterpiece, best quality 负向提示词 plaintext lowres, bad hands, missing fingers, extra fingers, mutated hands, poorly drawn face, ugly, deformed, blurry, text, watermark, modern clothes, western style, nsfw, extra limbs, bad proportions, cropped, jpeg artifacts正向提示词 plaintext Space station garden, floating alien flowers, glowing neon plants, starry universe background, distant planets, zero gravity, floating petals, futuristic metal corridor, cosmic light, sci-fi style, surreal scene, ultra detailed, cinematic lighting, masterpiece, best quality 负向提示词 plaintext lowres, blurry, text, watermark, ground, earth scenery, deformed plants, ugly, messy lines, worst quality, low quality, realistic photo, human crowd, extra limbs, bad perspective正向提示词 plaintext Retro European street corner cafe, vintage storefront, wrought iron window frame, potted roses on windowsill, warm afternoon sunlight, old street lamp, cobblestone road, lazy cat lying on the steps, retro tone, warm color palette, street view, detailed texture, masterpiece, best quality 负向提示词 plaintext lowres, blurry, text, watermark, logo, graffiti, modern billboard, deformed animals, ugly, disfigured, messy scene, bad proportions, cropped, jpeg artifacts, 3d anime distortion七、实验结论本实验完整掌握了千里云平台中 ComfyUI 节点式工作流的搭建、连接与参数配置流程熟悉了各核心节点的功能定位与协作逻辑四组基础文生图实验均成功生成图像整体流程运行稳定可靠。实验表明提示词撰写对最终生成效果具有决定性影响正向提示词需明确画面主题、风格、构图与细节描述负向提示词则可有效规避 AI 绘图中的各类常见缺陷。LoRA 模型能够以较低成本实现风格强化、角色定制与多风格融合单 LoRA 适用于单一风格优化多 LoRA 组合则可实现风格混搭配合分辨率调整能灵活平衡生成画质与运行效率。ComfyUI 的运行效率受分辨率、采样参数、LoRA 数量及节点布局等多因素共同影响采用标准分辨率、适中采样步数与精简节点结构可实现高效稳定出图。本次实验所用的千里云容器实例能够稳定支撑 ComfyUI 开展基础文生图、LoRA 增强绘图等操作可满足入门学习与常规创作的需求。实验不足与未来展望本实验仅采用单一采样器与固定调度器未对不同采样算法的生成效果差异进行对比也未探究不同 LoRA 权重对画面表现的影响因此实验内容存在一定局限性工作流节点也仍有优化空间。后续可开展拓展实验对比多款主流采样器的生成质量与运行效率测试不同 LoRA 权重的最优搭配方案也可引入 ControlNet 等进阶节点实现人物姿态、画面构图的精准控制还能利用队列功能完成批量生成进一步提升实操深度与创作效率。
基于ComfyUI的AI图像生成实验报告
一、实验引言1.实验目的熟悉千里云平台中ComfyUI节点式工作流的整体结构与核心节点作用掌握基础文生图工作流搭建、参数设置、正负提示词撰写以及图像生成完整流程。理解LoRA模型加载节点的功能学会单LoRA与多LoRA组合使用方式对比基础模型和LoRA增强模型的出图效果。同时探究采样器、采样步数、图像分辨率等参数对图像质量与程序运行效率造成的影响。二、实验环境与软硬件配置千里云容器实例ComfyUI在线运行环境访问地址https://4d8fef1c-8188fz.suanlihou.com三、实验原理ComfyUI采用节点串联的形式完成AI图像生成底层依托Stable Diffusion模型运行。基础大模型决定图像整体画风与画质基底空Latent图像会生成带有噪声的潜在图像空间也是最终画面的原始载体分辨率参数直接决定输出图片尺寸。CLIP文本编码能够将人为编写的正向、负向提示词转化为模型可识别的特征向量正向词用来定义画面内容、风格与细节负向词用于规避畸形、模糊、水印等缺陷。K采样器是核心推理模块依靠采样算法逐步去除噪声采样步数、CFG数值、采样器类型会同时影响画面细节和生成速度。VAE解码负责将处理后的潜在图像转化为常规像素图像最后由保存图像节点导出成品。LoRA属于轻量化微调模型依附基础大模型使用可强化特定角色、画风与装饰细节多个LoRA叠加还能实现多种风格融合。四、实验全流程步骤本次实验分为基础文生图、LoRA模型增强图像生成两大模块按照流程分步开展操作。一前期准备登录千里云容器实例通过指定网址进入ComfyUI操作界面。清空原有工作流依次添加并连接标准文生图所需节点顺序为Checkpoint加载器、空Latent图像、正负向CLIP文本编码、K采样器、VAE解码、保存图像。在模型加载节点中选定指定基础模型统一初始图像分辨率为512×512。二基础文生图实验第一组林间萤火木屋治愈森系风景第二组古风执扇江南少女国风人物第三组星际太空花园科幻幻想风第四组复古街角咖啡店欧式日常场景三LoRA模型增强实验在基础工作流的基础上新增LoRA加载器节点将其串联在基础模型加载节点五、实验结果与详细分析正向提示词 plaintext A small wooden house deep in the forest, twilight, countless fireflies flying around, dense green plants, old wooden windows, warm candlelight through the window, soft mist, peaceful atmosphere, natural scenery, soft focus, masterpiece, best quality, highly detailed 负向提示词通用防缺陷 plaintext lowres, bad anatomy, blurry, text, watermark, logo, signature, ugly, deformed, disfigured, extra objects, messy composition, worst quality, low quality, 3d render, cartoon distortion正向提示词 plaintext 1girl, traditional Chinese style, ancient costume, long hanfu, holding silk fan, long black hair tied with hairpin, standing by lotus pond, willow trees, rippling water, soft breeze, elegant posture, classical beauty, ink painting style, delicate facial features, vibrant colors, masterpiece, best quality 负向提示词 plaintext lowres, bad hands, missing fingers, extra fingers, mutated hands, poorly drawn face, ugly, deformed, blurry, text, watermark, modern clothes, western style, nsfw, extra limbs, bad proportions, cropped, jpeg artifacts正向提示词 plaintext Space station garden, floating alien flowers, glowing neon plants, starry universe background, distant planets, zero gravity, floating petals, futuristic metal corridor, cosmic light, sci-fi style, surreal scene, ultra detailed, cinematic lighting, masterpiece, best quality 负向提示词 plaintext lowres, blurry, text, watermark, ground, earth scenery, deformed plants, ugly, messy lines, worst quality, low quality, realistic photo, human crowd, extra limbs, bad perspective正向提示词 plaintext Retro European street corner cafe, vintage storefront, wrought iron window frame, potted roses on windowsill, warm afternoon sunlight, old street lamp, cobblestone road, lazy cat lying on the steps, retro tone, warm color palette, street view, detailed texture, masterpiece, best quality 负向提示词 plaintext lowres, blurry, text, watermark, logo, graffiti, modern billboard, deformed animals, ugly, disfigured, messy scene, bad proportions, cropped, jpeg artifacts, 3d anime distortion七、实验结论本实验完整掌握了千里云平台中 ComfyUI 节点式工作流的搭建、连接与参数配置流程熟悉了各核心节点的功能定位与协作逻辑四组基础文生图实验均成功生成图像整体流程运行稳定可靠。实验表明提示词撰写对最终生成效果具有决定性影响正向提示词需明确画面主题、风格、构图与细节描述负向提示词则可有效规避 AI 绘图中的各类常见缺陷。LoRA 模型能够以较低成本实现风格强化、角色定制与多风格融合单 LoRA 适用于单一风格优化多 LoRA 组合则可实现风格混搭配合分辨率调整能灵活平衡生成画质与运行效率。ComfyUI 的运行效率受分辨率、采样参数、LoRA 数量及节点布局等多因素共同影响采用标准分辨率、适中采样步数与精简节点结构可实现高效稳定出图。本次实验所用的千里云容器实例能够稳定支撑 ComfyUI 开展基础文生图、LoRA 增强绘图等操作可满足入门学习与常规创作的需求。实验不足与未来展望本实验仅采用单一采样器与固定调度器未对不同采样算法的生成效果差异进行对比也未探究不同 LoRA 权重对画面表现的影响因此实验内容存在一定局限性工作流节点也仍有优化空间。后续可开展拓展实验对比多款主流采样器的生成质量与运行效率测试不同 LoRA 权重的最优搭配方案也可引入 ControlNet 等进阶节点实现人物姿态、画面构图的精准控制还能利用队列功能完成批量生成进一步提升实操深度与创作效率。
