自研检测模型
使用数百万张真实照片和来自各大主流生成器的AI图片训练而成。它直接分析像素本身，因此即使所有水印和标签都被去除，也能识别AI图片。
超过500万用户正在使用
免费AI图片检测器
检测照片或艺术作品是否由AI生成或编辑。无论有无水印，我们的自研检测模型准确率都超过99%。
如何检查图片是否由AI生成
添加图片。
上传照片、将图片拖到页面上、按Ctrl+V粘贴，或使用图片链接。支持JPG、PNG、WebP、HEIC、AVIF和GIF格式，最大20 MB。
Phrasly开始检查。
扫描会立即开始。我们的自研检测模型逐像素分析图片，同时Phrasly会检查文件中的AI签名和SynthID水印。
查看结论。
查看检测结论、置信度分数，以及精确标出AI位置的热力图，所有检查项都会列在一旁。
永久免费，无需注册。您的图片绝不会被存储。
每次扫描的检查内容
自研模型为先，证据逐一核验
水印和元数据很容易被去除，因此不能作为唯一依据。Phrasly的自研检测模型直接分析图片本身，即使完全没有水印也能识别AI图片，就连只经过轻微AI编辑的照片也逃不过。如果文件中带有证据，我们也会一并检查。
SynthID水印
Google DeepMind推出的隐形水印，Google和OpenAI均在使用。一旦检测到，结论100%确定。
经过验证的AI签名
ChatGPT、Gemini、Adobe Firefly、Copilot和FLUX都会为生成的图片签名。我们会依据官方C2PA信任列表验证每一个签名。
细微AI编辑也能发现
换掉的背景、添加的物体、修过的面部。我们的模型能标记出仅部分经AI修改的照片，并精确标出位置。
截图同样适用
截图、重新上传的图片以及从社交媒体保存的副本都会丢失元数据。我们的模型并不依赖元数据。
可衡量的置信度
找到证据时为100%。否则显示我们模型自身的确定程度，让您清楚该结果的参考分量。
准确性
追求准确，而不只是快速
自研模型基于数百万张真实照片和AI图片训练，并额外核验文件内部的证据。
- 我们测试中的准确率，适用于有水印和无水印的AI图片。
- 99%+
- 发现SynthID水印或经过验证的AI签名时，结论完全确定。
- 100%
- 张图片被存储。每次扫描都在内存中完成，随后立即丢弃。
- 0
How We Test
We measure accuracy on real photos and AI images from every major generator, with and without watermarks. For the strongest result, upload the original file rather than a screenshot or a heavily compressed copy.
How We Compare to Other AI Image Detectors
Many free AI image detectors give one score from the pixels alone. Phrasly pairs a pixel model with SynthID and C2PA checks, shows a heatmap of the AI areas and never stores your image.
亲自试试
您能认出哪些是AI吗？
如今的AI图片已经非常逼真。猜猜其中哪些是AI生成的，再看看Phrasly在每张图片中发现了什么。
AI还是真实？
仔细观察，然后做出判断。
AI还是真实？
仔细观察，然后做出判断。
AI还是真实？
仔细观察，然后做出判断。
AI还是真实？
仔细观察，然后做出判断。
可检测所有主流AI图片生成器
每个月都有新的图片生成器问世，其中最出色的已经能生成几乎骗过所有人的图片。Phrasly能把它们全部识别出来。
有些图片在文件中带有证据，例如ChatGPT、Gemini和Adobe Firefly的签名。但很多图片没有，还有人会故意去除这些证据。我们的自研模型完全不需要这些：它直接分析像素本身，因此Midjourney、Stable Diffusion、FLUX等生成器的图片都能以同样的方式被识别。
我们可检测的AI图片生成器
ChatGPT和DALL-EGoogle GeminiNano BananaMidjourneyStable DiffusionFLUXAdobe FireflyGrokIdeogramMicrosoft CopilotRecraftLeonardo+ 更多
ChatGPT & DALL-E
ChatGPT images carry OpenAI's C2PA signature and, on newer models, a SynthID watermark. We check for both. When a screenshot strips them, the pixel model still catches the image.
Gemini & Nano Banana
Gemini and Nano Banana images carry Google's SynthID watermark and a signature that records it. We verify both, and fall back on pixel detection when they're gone.
Midjourney
Midjourney images rarely carry proof in the file. Our detection model reads the image itself, so Midjourney art and photos are caught without a watermark.
Stable Diffusion & FLUX
Open models run anywhere, often with no metadata. FLUX signs the images it makes; for everything else, our pixel model does the work.
Grok & Ideogram
Grok and Ideogram images are checked the same way: any signature or watermark first, then our model reads every pixel.
What you can check
Check AI Art, Photos, GIFs and Deepfakes
One AI image checker for every kind of picture you come across, from a viral photo to a piece of digital art.
AI Art Detector
Check digital art, illustrations, anime-style images and concept art submitted as hand-made work. The verdict shows how sure the model is, so you can decide what to do next.
AI Photo Detector
Camera-style AI photos are the hardest to spot by eye. Upload a photo to see if it was generated, with a heatmap of any AI areas.
AI GIF Detector
Upload a GIF up to 20 MB in the same box as any photo, and get the same verdict, confidence score and heatmap.
Deepfakes and Face Swaps
A swapped face or an AI-retouched portrait is flagged as Partially AI, with the changed areas highlighted. Phrasly checks still images, not video.
Partially AI-Edited Photos
Generative fill, a new background or an added object: when only part of a real photo was changed with AI, the heatmap marks exactly where.
Guide
How to Tell If a Picture Is AI-Generated
For a long time you could tell if a picture was AI by looking closely. Older image generators left clear signs, and most guides still list them (see the checklist).
New models like Nano Banana, Midjourney and FLUX have fixed most of these. Hands look right, text is readable and the lighting holds together, so a careful look is no longer enough to answer "is this photo AI?". Screenshots and re-uploads also wipe the file details that used to give an image away.
The reliable way to check if an image is AI-generated is an AI photo detector that reads the pixels and the file at the same time. Phrasly does both and shows its evidence, so you see why an image was called AI, not just a score.
Signs an image may be AI-generated
- Hands and fingers Extra, missing or merged fingers.
- Text in the image Signs, labels and logos that turn into nonsense letters.
- Lighting and shadows Shadows falling the wrong way, reflections that don't match.
- Backgrounds Faces, windows or patterns that melt into each other.
- Skin and texture Skin that looks too smooth, plastic or airbrushed.
- File details No camera data, or a C2PA signature naming an AI tool.
谁在使用Phrasly的AI图片检测器
从新闻编辑部到课堂，在信任一张图片之前，先检查一下。
记者和事实核查人员
在发布前核实热门照片或消息来源提供的图片，并获得可引用的证据。
教师和学校
检查作为原创作品提交的艺术作品、摄影作业和项目图片。
交易平台和约会应用
在AI生成的头像和商品图片触达用户之前将其拦截。
品牌和创作者
确认您获得授权、发布或转发的内容是真实的。
常见问题
常见问题解答
是的，永久免费。无需账户、无需注册，也无需信用卡。每个结果都包含检测结论及其理由；只要图片可以判定，还会提供置信度分数和热力图。