Defect Creator Studio

Problem it solves: 你需要用于罕见缺陷的训练数据，但等待它们自然发生需要数周或数月。

What it does: 通过从一张良好图像和一个简单的英文描述生成照片级真实的合成缺陷图像。上传一张良品图片，标注缺陷应出现的位置，描述它（"焊缝处的细小裂纹"），即可在数秒内获得数十张逼真的训练图像。

Getting started

1. Upload 一张干净、光线充足的部件图片。将图片拖放到画布，或点击 Browse Files。该工具支持高达 8K 分辨率，所有图片都在浏览器本地存储。
2. Choose your camera target. 在 OV10i/OV20i (1408 x 1080，4:3) 与 OV80i (3840 x 2160，16:9) 之间切换。这些与 OV 相机的实际传感器输出相匹配。如果图像尺寸较大，工具会自动缩放。如果长宽比不匹配，裁剪覆盖层将允许你选择要保留的部分。
3. Select a defect type. 上传后，AI 会分析部件的材料、形状和表面处理，并建议相关的缺陷类型。共提供 16 种内置缺陷类型：Scratch、Dent、Chip、Stain、Crack、Corrosion、Porosity、Weld Defect、Discoloration、Burr、Delamination、Warping、Contamination、Missing Material、Inclusion 和 Oxidation。你也可以使用自己的描述添加 Custom Defects。
4. Mark the defect region. 使用标注工具在图像上精确绘制缺陷应出现的位置。
5. Generate. 点击紫色的 Generate 按钮。AI 会在你标记的区域内创建一个逼真的缺陷。

Annotation tools

左侧工具栏提供七种工具，用于精确放置缺陷：

Shortcut	Tool	Best for
C	Circle Marker	圆形或点状缺陷，如凹坑、气泡、局部变色
M	Rectangle Select	带状图案，如边缘的划痕或面板级翘曲
L	Lasso	不规则或有机形状，如裂纹、溢出、复杂断裂模式
G	Move/Resize	绘制后重新定位和调整标注大小
E	Eraser	删除标注
H	Pan	在图像中导航（按住 Space 也可临时使用）
Z	Zoom	放大至 800% 以实现像素级精度

额外的快捷键：Ctrl+Z 撤销，Ctrl+Shift+Z 重做，Del 删除最后一个标注。

Annotation toolbar

Defect type panel with camera selector and Generate button

Tight regions produce better results

将每个标注视为一个精确的指令窗口。AI 仅在你选定的区域内生成缺陷。如果你的区域紧贴目标缺陷区域，模型可以更精准地聚焦。包含不相关背景的宽松区域可能导致模型将缺陷影响扩散到额外区域。

编写有效的缺陷描述

AI 将您的缺陷名称解读为自然语言指令。具体性很重要。

Quality	Example	Why
Good	"Light horizontal transparent scratches on glossy plastic"	Includes morphology, direction, surface type, and visual behavior
Good	"Fine radial crack near molded corner"	Specific about shape, location, and material context
Bad	"scratch"	Too vague for the model to produce anything useful
Bad	"damage"	No morphology, no surface context

请先从 AI 建议的缺陷类型开始。它们在 UI 中显示得很短，但在内部每个类型都映射到更丰富的语义描述，帮助模型产生更准确的结果。只有在目标未被这些建议覆盖时，才添加自定义缺陷。

从图像捕获（参考缺陷）

这是最强大的功能之一。使用 Browse & Capture 选择已上传的图像，标记实际缺陷区域，并将其保存为可重复使用的参考模板。这样可以将一个实际缺陷从一张图像转移到其他图像。

• Naming is critical. The crop includes both the defect and surrounding background. The AI uses your name to understand which part is the actual defect. "Bent pin on connector" gives the model a semantic anchor, while "defect" leaves it guessing.
• Transfer scenarios: Same part in a different position, similar parts on the same line, or even different manufacturing lines when the defect concept still makes sense (e.g., a scratch on one plastic housing transfers well to another plastic housing).
• Severity, Rotation, and Size sliders (reference defects only): Without variation, every transferred defect would be an exact copy. Rotation changes the angle and orientation. Size controls spatial coverage. Severity controls how visually prominent the defect appears.

严重性控制

The severity slider controls how visually prominent the defect appears. The default is intentionally subtle because most real manufacturing defects are subtle. Even at 50%, defects remain quite faint. At 1%, defects are nearly invisible to humans, but OV vision cameras can still detect them. This is one of the key advantages of machine vision: the camera never gets tired and can pick up patterns below the threshold of human perception.

风格变体模式

Located next to the Change Image button. After uploading an image, the AI analyzes the part type and suggests style modifications common in manufacturing, like changing a metal finish from brushed to polished, switching plastic color, or altering coating type. You can also add your own custom style variations.

Two valid sequences:

• Restyle first, then add defects. Useful for generating training data for a new product line or colorway before it physically exists.
• Add defects first, then restyle. See how the same defect looks across material, finish, and color variants.

Part geometry, orientation, layout, and composition remain identical. Only the targeted style attribute changes. This is especially valuable because the defect context (location, shape, severity) stays stable across variants, so your model learns to detect the defect itself rather than a specific combination of defect plus surface appearance.

区域限制

• 最多有 9 个常规缺陷区域（内置或基于文本的自定义）
• 最多有 4 个参考图像传输缺陷
• 区域数量通常越少，结果越干净。请从一个精确的区域开始，验证质量，然后逐步增加。

队列与吞吐量

The Studio runs up to 3 generations in parallel. Additional requests queue automatically and execute as slots become available. Use 3-run bursts for rapid A/B/C comparison: submit the same configuration three times, compare results side by side, and pick the best output. For volume, queue 10 or more jobs and let them process while you continue other work.

• The Studio 最多可同时进行 3 个生成任务。附加请求会自动排队，并在槽位可用时执行。为快速进行 A/B/C 比较，请使用 3 轮运行：对同一配置提交 3 次，并排比较结果，择优输出。对于大批量，请将 10 个或以上的作业加入队列，让它们在您继续其他工作时进行处理。

Compare, Library, and Downloads

• Compare: Toggle between the baseline and generated image. For subtle defects, rapid toggle/flicker-style switching works best. The human visual system detects change through motion far better than through static side-by-side comparison.

• Image Library: The bottom tray provides a scrollable filmstrip for quick visual scanning. The expanded view adds full dataset management: search, sort, multi-select (Ctrl/Shift-click), download individual images, create ZIP archives for training pipelines, and delete.

• Annotation persistence: When you switch between images, all annotations are automatically preserved and restored. Navigate away to review another image, then come back to find all your regions exactly where you left them.

• Compare: 在基线图像与生成图像之间切换。对于微小缺陷，快速切换/闪烁式切换效果最佳。人眼视觉系统通过运动检测变化，远比静态并排对比更易察觉。

• Image Library: 底部托盘提供一个可滚动的胶片条，便于快速进行视觉浏览。扩展视图增加了完整的数据集管理功能：搜索、排序、多选（Ctrl/Shift-click）、下载单张图像、为训练管道创建 ZIP 存档，以及删除。

• Annotation persistence: 当在图像之间切换时，所有注释都会自动保存并在返回时恢复。切换到其他图像进行查看后再返回，您将看到所有区域仍然保留在离开时的位置。

推荐工作流程

1. 上传一张干净的基线图像，确保其适配 OV80i（3840 x 2160）的取景。良好的照明、正确对焦，以及背景杂乱的最小化都有助于提高结果。
2. 以 AI 提供的缺陷建议为起点。它们是针对您的零件类型进行校准的。仅在目标缺陷缺失时添加自定义命名。
3. 放置一个精确、紧凑的区域并生成初步结果。单一区域运行是最可靠的基线。
4. 每次生成后都要积极使用 Compare。对于微小缺陷，快速切换能够让连微小的差异在视觉上突出。
5. 根据所见调整缺陷严重性、区域紧密度和措辞。重复操作直到质量稳定，然后扩展到多区域和排队量。

Start with real data, accelerate with synthetic

最佳做法：先用初始的 3-5 张真实图像进行训练，找出 AI 的挑战点，然后使用 Defect Studio 为这些特定故障模式生成有针对性的合成示例。真实数据提供基线，合成数据填补空白。

Synthetic data supplements real data, it does not replace it

Defect Studio 的图像对于填补训练集中的空白非常有用，但它们绝不能成为唯一的训练数据来源。请始终用真实生产图像来验证模型性能。

See it in action

Where it fits in the workflow: You'll use Defect Studio during Step 4: Train Your AI Model to build training data faster.