REDDIZ — Vision Dataset Builder & Reddit Annotator

A lightweight, zero-dependency annotation tool I built to pull image feeds directly from Reddit (including NSFW and multi-subreddit searches), crop them to 9:16 vertical preview framing, annotate them quickly, and export clean training datasets for computer vision models, LoRAs, and vision-language models.

I made this because manual dataset scraping, cropping, and repetitive labeling is usually tedious. REDDIZ streamlines the process with auto-propagating prompts, fast keyboard shortcuts, and built-in ZIP generation without needing external Python libraries or complex setup.

• Multi-Subreddit Fetching: Enter single or multiple subreddits separated by commas or spaces (e.g. cats, EarthPorn, streetphotography).
• NSFW & Rate-Limit Handling: Uses age-verification session headers and exponential backoff retry logic to bypass Reddit's aggressive 429 throttling and hotlink blocks.
• 9:16 Center Crop Viewport: Automatically previews and crops images from the center into a modern 9:16 aspect ratio.
• Vertical Gallery Strip: A scrollable vertical carousel on the left that lets you flick through thumbnails with the up and down arrow keys.
• Fast Annotation Flow:
  • Prompt/caption and tags carry over to the next image automatically so you don't re-type identical tokens.
  • Bounding boxes start clean on each new image.
  • One-click × CLEAR buttons for caption text and bounding boxes.
  • ⊘ IGNORE option (or hit X / Delete) to exclude bad or irrelevant images from the final export.
• Direct Multi-Format Export:
  • YOLOv8 / YOLOv11: Normalized labels/*.txt, classes.txt, and data.yaml.
  • COCO JSON: Industry-standard coco_annotations.json.
  • Stable Diffusion / Flux / LoRA: Individual .txt prompt files paired alongside each image.
  • Vision-Language (VLM): dataset.jsonl formatted for multimodal fine-tuning.

Make sure you have Node.js installed (v16+ recommended). No npm install needed—everything runs on native standard Node modules.

1. Clone the repo:

   git clone <your-repo-url>
   cd RedditDataTrainer

2. Start the server:

   node server.js

3. Open the browser:

   http://localhost:3000
   

Type any subreddit into the search bar at the top, select your sort preference (NEW, HOT, TOP), hit FETCH MEDIA, and start annotating.

RedditDataTrainer/
├── server.js          # Native HTTP server, Reddit RSS parser, image proxy & ZIP packager
├── simple-zip.js      # Zero-dependency streaming Deflate ZIP generator (Node zlib)
├── gemini.js          # Google Gemini Vision integration
├── openai.js          # OpenAI GPT-4o Vision integration
├── claude.js          # Anthropic Claude 3.5 Sonnet Vision integration
├── groq.js            # Groq Llama 3.2 Vision integration
├── huggingface.js     # Hugging Face Inference API integration
├── vercel.json        # Vercel serverless deployment config
├── package.json       # Project manifest
├── README.md          # Project documentation
└── public/
    ├── index.html     # Editorial 3-column UI layout
    ├── app.css        # Modern typography & brutalist styling
    ├── app.js         # Canvas bounding boxes, state inheritance, and hotkeys
    └── favicon.svg    # App icon

When you click DONE & DOWNLOAD ZIP, your download includes:

reddit_dataset.zip
├── images/
│   ├── 0001_abc123.jpg
│   └── 0002_def456.png
├── labels/                # (YOLO format)
│   ├── 0001_abc123.txt
│   └── 0002_def456.txt
├── captions/              # (Stable Diffusion / LoRA txt files)
│   ├── 0001_abc123.txt
│   └── 0002_def456.txt
├── data.yaml              # YOLO configuration
├── classes.txt            # List of bounding box classes
├── dataset.jsonl          # Multimodal dataset
└── manifest.json          # Complete metadata log with source URLs

• Ignored images are skipped automatically during export so they won't clutter your training folders.
• If Reddit returns a temporary 429 rate limit, the backend waits and retries automatically. Giving it a few seconds between heavy multi-subreddit fetches helps keep queries reliable.