ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1.75em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">
ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1.75em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">最近全球刷屏Nano Banana，应该没有人还没用过吧？！ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1.75em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">它不仅能根据一句话描述，就生成栩栩如生的手办图片，还能根据用户的描述，对图片进行精细化编辑，就连速度，也快得出奇。ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">（指令：为马斯克换帽子和裙子。可以看到右图所有要素除了稍微漏了裤边之外，融合还算不错，甚至还贴心考虑到了穿裙子需要把短袖扎进去的细节，整个生成耗时仅16.0s。）ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1.75em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">可以说，作为当下最优秀的生图模型，Nano Banana在一致性以及精细度上，已经做到了符合企业级生产标准。ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1.75em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">比如，我们服务的一家集抽卡、换装于一体的娱乐公司，最近正在开发一个功能，通过接入Nano Banana，实现用户上传照片后，可以自由从素材库中选择喜欢的配饰、道具进行换装打扮。ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1.75em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">一些电商客户，也在考虑用AI为模特换装、换发型、换配饰，实现一次拍摄，永久使用。ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1.75em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">从这两个案例中，我们不难发现，对很多企业级用户来说，仅有一个不错的生图模型还不够。他们还需要足够聪明的检索，从海量历史素材中，找到最合适的服装、配饰、以及其他人物元素。ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1.75em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">也就是说，用户需要的是一个集向量数据库+图片生成模型的多模态RAG系统。ingFang SC", system-ui, -apple-system, "system-ui", "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 17px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: 0.544px;orphans: 2;text-align: justify;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;white-space: normal;text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;background-color: rgb(255, 255, 255);line-height: 1.75em;visibility: visible;box-sizing: border-box !important;overflow-wrap: break-word !important;">那么，如何用Nano Banana+Milvus向量数据库，搭建这样一套生产级的多模态RAG系统？本文将给出手把手教程。01搭建以文搜图系统对一些快消品公司以及游戏娱乐公司来说，用AI生图最大的问题不在于生成，而在于历史素材过多，但这些素材都是图像、音频、视频这样的非结构化数据。常规方法下，我们无法对其进行精准的检索召回。 因此，在这一步，我们需要先搭建一个完善的以文搜图系统。 我们可以使用 CLIP 模型将图像和文本转为向量，然后将向量存储到 Milvus 数据库，最后通过 Milvus 向量数据库进行高效的相似性搜索（用户通过文字描述即可搜索图片，并返回 top 3 结果）。 以下是具体教程： 安装依赖包
#安装必要的包%pipinstall--upgradepymilvuspillowmatplotlib%pipinstallgit+https://github.com/openai/CLIP.git 导入必要的库
importosimportclipimporttorchfromPILimportImageimportmatplotlib.pyplotaspltfrompymilvusimportMilvusClientfromglobimportglobimportmathprint("所有库导入成功！")初始化Milvus客户端
#初始化Milvus客户端milvus_client=MilvusClient(uri="http://localhost:19530",token="root:Miluvs")print("Milvus客户端初始化成功！")加载CLIP模型
#加载CLIP模型model_name="ViT-B/32"device="cuda"iftorch.cuda.is_available()else"cpu"model,preprocess=clip.load(model_name,device=device)model.eval()print(f"CLIP模型'{model_name}'加载成功，运行设备:{device}")print(f"模型输入分辨率:{model.visual.input_resolution}")print(f"上下文长度:{model.context_length}")print(f"词汇表大小:{model.vocab_size}")定义特征提取函数
defencode_image(image_path):"""将图像编码为归一化的特征向量"""try:image=preprocess(Image.open(image_path)).unsqueeze(0).to(device)withtorch.no_grad():image_features=model.encode_image(image)image_features/=image_features.norm(dim=-1,keepdim=True)#归一化returnimage_features.squeeze().cpu().tolist()exceptExceptionase:print(f"处理图像{image_path}时出错:{e}")returnNonedefencode_text(text):"""将文本编码为归一化的特征向量"""text_tokens=clip.tokenize([text]).to(device)withtorch.no_grad():text_features=model.encode_text(text_tokens)text_features/=text_features.norm(dim=-1,keepdim=True)#归一化returntext_features.squeeze().cpu().tolist()print("特征提取函数定义成功！")创建Milvus集合
collection_name="production_image_collection"#如果集合已存在，删除它ifmilvus_client.has_collection(collection_name):milvus_client.drop_collection(collection_name)print(f"已删除现有集合:{collection_name}")#创建新集合milvus_client.create_collection(collection_name=collection_name,dimension=512,#CLIPViT-B/32的嵌入维度auto_id=True,#自动生成IDenable_dynamic_field=True,#启用动态字段metric_type="COSINE"#使用余弦相似度)print(f"集合'{collection_name}'创建成功！")print(f"集合信息:{milvus_client.describe_collection(collection_name)}")处理并插入图像
#设置图像目录路径image_dir="./production_image"raw_data=[]#获取所有支持的图像格式image_extensions=['*.jpg','*.jpeg','*.png','*.JPEG','*.JPG','*.PNG']image_paths=[]forextinimage_extensions:image_paths.extend(glob(os.path.join(image_dir,ext)))print(f"在{image_dir}中找到{len(image_paths)}张图像")#处理图像并生成嵌入successful_count=0fori,image_pathinenumerate(image_paths):print(f"处理进度:{i+1}/{len(image_paths)}-{os.path.basename(image_path)}")image_embedding=encode_image(image_path)ifimage_embeddingisnotNone:image_dict={"vector":image_embedding,"filepath":image_path,"filename" s.path.basename(image_path)}raw_data.append(image_dict)successful_count+=1print(f"成功处理{successful_count}张图像")将数据插入Milvus
#将数据插入Milvusifraw_data:print("正在将数据插入Milvus...")insert_result=milvus_client.insert(collection_name=collection_name,data=raw_data)print(f"成功插入{insert_result['insert_count']}张图像到Milvus")print(f"插入的ID示例:{insert_result['ids'][:5]}...")#显示前5个IDelse:print("没有成功处理的图像数据可插入")定义搜索和可视化函数
defsearch_images_by_text(query_text,top_k=3):"""根据文本查询搜索图像"""print(f"搜索查询:'{query_text}'")#编码查询文本query_embedding=encode_text(query_text)#在Milvus中搜索search_results=milvus_client.search(collection_name=collection_name,data=[query_embedding],limit=top_k,output_fields=["filepath","filename"])returnsearch_results[0]defvisualize_search_results(query_text,results):"""可视化搜索结果"""num_images=len(results)ifnum_images==0:print("没有找到匹配的图像")return#创建子图fig,axes=plt.subplots(1,num_images,figsize=(5*num_images,5))fig.suptitle(f'搜索结果:"{query_text}"(Top{num_images})',fontsize=16,fontweight='bold')#处理单个图像的情况ifnum_images==1:axes=[axes]#显示图像fori,resultinenumerate(results):try:img_path=result['entity']['filepath']filename=result['entity']['filename']score=result['distance']#加载并显示图像img=Image.open(img_path)axes[i].imshow(img)axes[i].set_title(f"{filename}\n相似度:{score:.3f}",fontsize=10)axes[i].axis('off')print(f"{i+1}.文件:{filename},相似度分数:{score:.4f}")exceptExceptionase:axes[i].text(0.5,0.5,f'加载图像出错\n{str(e)}',ha='center',va='center',transform=axes[i].transAxes)axes[i].axis('off')plt.tight_layout()plt.show()print("搜索和可视化函数定义成功！")10.执行文本搜索图像#示例搜索1query1="agoldenwatch"results1=search_images_by_text(query1,top_k=3)visualize_search_results(query1,results1) 02用Nano-banana为品牌创作宣传图安装Google SDK%pipinstallgoogle-generativeai%pipinstallrequestsprint("GoogleGenerativeAISDK安装完成！")配置GeminiAPIimportgoogle.generativeaiasgenaifromPILimportImagefromioimportBytesIOgenai.configure(api_key="<your_api_key>") 生成新图像prompt=("AnEuropeanmalemodelwearingasuit,carryingagoldwatch.")image=Image.open("/path/to/image/watch.jpg")model=genai.GenerativeModel('gemini-2.5-flash-image-preview')response=model.generate_content([prompt,image])forpartinresponse.candidates[0].content.parts:ifpart.textisnotNone:print(part.text)elifpart.inline_dataisnotNone:image=Image.open(BytesIO(part.inline_data.data))image.save("generated_image.png")image.show()
03 效果演示除了上文展示的场景，我们不妨把脑洞再放大，比如某品牌发布了很多新品，又不想重新找模特进行拍摄，就可以靠Nano-banana直接搞定宣传图 Prompt: A model is wearing these products on the beach 除了简单场景，我们也能实现一些天马行空的，随意叠加的场景，物品，人物。 Prompt: A model is posing and leaning against a blue convertible sports car. She is wearing a halter top dress and the accompanying accessories. She is adorned with a diamond necklace and a blue watch, wearing high heels on her feet and holding a labubu pendant in her hand. 最后就是最常见的手办原型制作，比如我们最近想做一些可爱的手办，就可以先让nano banana先来代劳。 Prompt: Use the nano-banana model to create a 1/7 scale commercialized figure of thecharacter in the illustration, in a realistic styie and environment.Place the figure on a computer desk, using a circular transparent acrylic base without any text.On the computer screen, display the ZBrush modeling process of the figure.Next to the computer screen, place a BANDAl-style toy packaging box printedwith the original artwork. 整体测评下来，我们发现Nano-banana完全担得起当下最强AI生图模型的称号。不仅做到了高一致性、微调的可控性，甚至能兼顾到水中倒影，产品模型图与实物图、包装logo图需要统一的魔鬼细节。 但Nano-banana并非完美无缺，在一些非常专业的场景中，依然会出现复杂指令理解失误或者光影不科学的小问题。我们可以考虑在提示词外，给AI一些想要的风格参考，或者对光源来源、光影效果给出具体描述，这样可以事半功倍哦。 |
