Dataset Preparation Guide

Image Editing Training (AI Toolkit)

This trainer endpoint fine-tunes an image editing diffusion model (e.g. inpainting, instruction-based editing) using ai-toolkit in the background.

Image editing training requires paired data: an original image, an edited target image, and optionally a mask.

1. Dataset Format (Required)

Your dataset must be a single folder containing image pairs. Each edit example is identified by a shared base filename.

Folder Structure

dataset/
├── edit_001_input.png
├── edit_001_target.png
├── edit_001.txt
├── edit_002_input.jpg
├── edit_002_target.jpg
├── edit_002.txt

Optional (for inpainting / masked editing)

dataset/
├── edit_003_input.png
├── edit_003_target.png
├── edit_003_mask.png
├── edit_003.txt

Naming Rules

• _input → original image
• _target → edited result (what the model should produce)
• _mask → optional binary mask (white = editable area)
• .txt → edit instruction

All files must share the same base name (e.g. edit_003).

2. Instruction Files (.txt)

Each .txt file contains the editing instruction describing how to transform the input image into the target image.

Example

edit_001.txt

replace the background with a snowy mountain landscape

Instruction Guidelines

• Describe only the change
• Do not restate the full image description
• Write instructions as imperative edits
• One instruction per file (multi-line allowed)

3. Mask Files (Optional but Recommended)

For tasks such as: - inpainting - object replacement - localized edits

you can include a mask image.

Mask Rules

• Same resolution as input image
• White (255) = editable region
• Black (0) = frozen region
• Format: .png recommended

If no mask is provided, the model assumes global editing.

4. Image Requirements

ai-toolkit automatically handles resizing and bucketing.

Recommended

• Resolution: ≥ 512×512
• Formats: .jpg, .png, .webp
• Input and target images must:
• have the same resolution
• be pixel-aligned
• differ only in the edited regions

Avoid

• Misaligned input/target pairs
• Style changes unrelated to the instruction
• Multiple edits per example

5. Dataset Size Recommendations

Editing models require more data than concept fine-tuning.

6. What Not to Include

• Missing input/target pairs
• Incorrect filename suffixes
• Non-binary masks
• Captions instead of edit instructions
• Copyrighted or unlicensed images

7. Uploading the Dataset

Once your dataset folder is ready:

1. Upload the dataset folder to the trainer endpoint
2. Select the editing mode (global / masked)
3. Start training

The trainer validates all image pairs before launching training.

8. Minimal Example Dataset

my_edit_dataset/
├── room_01_input.png
├── room_01_target.png
├── room_01_mask.png
├── room_01.txt → replace the sofa with a wooden table
├── room_02_input.jpg
├── room_02_target.jpg
├── room_02.txt → change the lighting to warm sunset light

Need Help?

If you are unsure whether your dataset is valid or want feedback on edit instructions, reach out before starting training — most training failures come from dataset issues.