ArtiAgent Web UI — Synthetic Defect Generation

1 Product & Defect Type

Product Name (Object Class)

This becomes the checkpoint folder name. Use lowercase with underscores.

Defect Type Name

The specific defect category to learn (subfolder name).

📋 Quick Guide

Enter product & defect names
Upload ≥1 defect images
Draw rectangle masks on defects
Add domain instructions (optional)
Start training & wait
Switch to "Generate Defects" tab

1 Product Verification (Detect Similar Product)

Not sure if this product model is trained? Upload 1 clean image to auto-detect and select the trained model. (Make sure the image is aligned, not random orientations)

🔍

Click to upload clean product image

PNG/JPG/BMP/TIF accepted

2 Synthetic Defect Generation (DefectFill)

Trained Product (Object Class) *

Trained checkpoints auto-populate here. Train a product first.

Defect Type (Optional)

Leave empty to let the VLM choose the most plausible defect.

Defect Area Detection Type *

Select "Auto" to let the VLM decide, or manually choose to guide the perception module.

Product Description *

Be specific. This drives VLM defect planning and placement.

Number of Defects per Image *

Guidance Scale *

Denoising Steps

Image Size

Input Mode

Single Image Folder

Clean Product Image *

Upload a clean (non-defective) image of the product.

Device

🖼️ Generation Results

Generated images will appear here.
Fill the form and click Generate.

⚙️ DefectDiffu Model Configuration

DefectDiffu Checkpoint Path *

Absolute path to the trained DefectDiffu checkpoint (.pt file).

VAE Path *

Path to Stable Diffusion VAE (HuggingFace model name or local path).

VLM Model

Local VLM model for defect planning (e.g., gemma3:12b).

✨ Generate Defects (DefectDiffu)

Product Description *

Be specific. This drives VLM defect planning and DefectDiffu generation.

Target Defect Type (Optional)

Leave empty to let the VLM choose the most plausible defect type.

Max Defects per Image *

Denoising Steps

Image Size

Device

Input Mode

Single Image Folder

Clean Product Image *

Upload a clean (non-defective) image of the product.

🖼️ DefectDiffu Generation Results

Generated images will appear here.
Configure model paths, fill the form, and click Generate.

📋 DefectDiffu Guide

Set DefectDiffu checkpoint & VAE paths
Enter product description
Optionally specify a target defect type
Upload clean product image(s)
Click Generate — no training needed!

Note: DefectDiffu uses text-guided diffusion with three disentangled prompts (c_p, c_d, c_f) and does not require few-shot training like DefectFill.

🔧 ArtiAgent Web UI Checking CUDA...

1 Product & Defect Type

2 Upload Good (Clean) Images

3 Upload Defect Images

3.5 Image Alignment

4 Draw Defect Masks

5 Domain-Specific Instructions

Example for Triac:

6 Training Configuration

📊 Training Monitor

📋 Quick Guide

1 Product Verification (Detect Similar Product)

2 Synthetic Defect Generation (DefectFill)

🖼️ Generation Results

⚙️ DefectDiffu Model Configuration

✨ Generate Defects (DefectDiffu)

🖼️ DefectDiffu Generation Results

📋 DefectDiffu Guide

1 Product & Defect Type

2 Upload Good (Clean) Images

3 Upload Defect Images

3.5 Image Alignment

4 Draw Defect Masks

5 Domain-Specific Instructions

Example for Triac:

6 Training Configuration

📊 Training Monitor

📋 Quick Guide

1 Product Verification (Detect Similar Product)

2 Synthetic Defect Generation (DefectFill)

🖼️ Generation Results

⚙️ DefectDiffu Model Configuration

✨ Generate Defects (DefectDiffu)

🖼️ DefectDiffu Generation Results

📋 DefectDiffu Guide

⚠️ Product Already Exists