---
language:
- en
license: mit
tags:
- image-dataset
- lora-training
- sdxl
- cartoon
- comic
- black-and-white
- animation
- captioned
size_categories:
- 1K<n<10K
---

# Spy vs Spy Training Dataset

Captioned image dataset for training SDXL LoRAs in the **Spy vs Spy** art style. Contains black-and-white animation frames and comic panels paired with descriptive `.txt` captions formatted for kohya_ss DreamBooth/LoRA training.

## Versions

| Version | Images | Source | Captioning Method |
|---------|--------|--------|-------------------|
| **v1** | 36 | MAD Magazine comic panels | Manual descriptions expanded by AI |
| **v2** | 220 | v1 panels + MadTV animation frames | Voice-guided Claude Vision + hallucination cleanup |
| **v3** | 861 | MadTV animated series (DVD rips) | Gemini video scene descriptions |

## Structure

Each version is in its own directory. Images are organized in kohya_ss format with repeat counts in the folder name:

```
v1/
  11_spyvspy/
    01.png
    01.txt
    02.png
    02.txt
    ...
v2/
  2_spyvspy/
    frame_001.png
    frame_001.txt
    ...
v3/
  1_spyvspy/
    s1_e01_a_s01.jpg
    s1_e01_a_s01.txt
    ...
```

The folder prefix number (e.g. `11_spyvspy`) is the repeat count — how many times each image is seen per epoch during training. Lower image counts use higher repeats to reach a similar number of training steps per epoch.

## Caption Format

Each `.txt` file contains comma-separated tags starting with the trigger word `spyvspy`:

```
spyvspy, full body view, white spy planting bomb under table, black spy sneaking up behind with mallet, indoor dining room setting, both wearing fedora hats and trenchcoats with long pointed beak noses and black sclera eyes
```

### Caption Structure

```
{trigger_words}, {framing}, {scene tags}, {appearance suffix}
```

- **Trigger words:** `spyvspy` (style), `white_spy` / `black_spy` (characters)
- **Framing:** full body view, close-up, upper body view, extreme close-up, etc.
- **Scene tags:** character actions, objects, environment
- **Appearance suffix:** physical description of the characters (fedora, trenchcoat, beak nose, black sclera eyes)

## Image Preparation

All images were:
1. Upscaled 4x with [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN) using the `4x-AnimeSharp` model
2. Downscaled to 1024px on the longest side
3. Converted to grayscale to match the black-and-white art style

## Data Collection Pipeline

### v1
36 comic panels manually collected from MAD Magazine. Captions written by hand and expanded with AI assistance.

### v2
184 frames hand-selected from ~44,800 extracted animation frames across 2 seasons of MadTV's Spy vs Spy. During selection, a voice note was recorded describing each frame. The voice note was transcribed and sent with the frame to Claude Vision for detailed captioning. A second pass compared the AI caption against the voice note to remove hallucinated objects.

Combined with the 36 v1 comic panels for a total of 220 images.

### v3
Full episodes sent to Gemini for video-level scene analysis. Frames extracted at Gemini-identified timestamps using FFmpeg with yadif deinterlacing and mpdecimate duplicate removal. Each frame reviewed in a custom PySide6 desktop app for quality selection and caption editing.

861 frames across 54 episodes, 2 seasons.

## Links

- **Trained LoRA checkpoints:** [camdenbalberg/spy-vs-spy-lora](https://huggingface.co/camdenbalberg/spy-vs-spy-lora)
- **Training pipeline & code:** [github.com/camdenbalberg/spy-vs-spy-lora](https://github.com/camdenbalberg/spy-vs-spy-lora)