ModelCollapser

This script uses OpenAI's gpt-image-1 model to perform iterative image edits: it takes a base image and a text prompt, generates a new image, then uses that output as the next input, repeating for N iterations.
Usually it leads to model collapse.

What is Model Collapse?

Model collapse is a phenomenon that occurs when an AI model is repeatedly asked to generate variations of its own outputs. Over multiple iterations, the model tends to:

1. Lose Diversity: The model starts producing increasingly similar outputs, eventually converging to a limited set of patterns or styles.
2. Forget Original Features: Details from the original image gradually fade away as the model focuses on the most prominent features.
3. Amplify Biases: The model's inherent biases become more pronounced with each iteration.
4. Create Artifacts: Strange patterns or distortions may emerge as the model struggles to maintain coherence.

This tool demonstrates model collapse by:

• Taking a base image and applying the same prompt repeatedly
• Showing how the image evolves (or devolves) through multiple generations
• Highlighting the limitations of current AI image generation models

The results can be both fascinating and concerning, providing insights into the stability and reliability of AI image generation systems.

---

🚀 Features

• Iterative edits: produce a chain of variants by feeding each output back as input.
• Auto-retry: optional --retries flag to automatically retry failed API calls. Useful for when image or prompt is not passing consistently OpenAI filters - this sometimes happen, and image is generated only after some retries.
• Verbose output: prints progress and file paths as it runs.
• Unique output folders: each run creates a new UUID-named directory to avoid collisions.
• Gradio Web UI: run as a web app for easy image uploads and prompt entry.
• Visualization Tools:
  • Animated GIF: automatically generates a GIF showing the progression of changes
  • Image Gallery: view all intermediate steps in a scrollable grid layout
  • Final Result: see the end result of the generation chain
• Docker Support: run the app in a containerized environment.

📸 Screenshots

The Gradio web interface makes it easy to upload images and generate variations.

Watch how the image evolves through multiple iterations in this animated GIF.

---

📋 Prerequisites

• Python 3.7+
• OpenAI Python package
• [Optional] Docker (for containerized usage)

Install dependencies:

pip install -r requirements.txt

---

🔧 Setup

1. Clone or download this repository, and ensure main.py and app.py are in your working dir.

2. Set your API key:

   export OPENAI_API_KEY="YOUR_OPENAI_API_KEY"

---

⚙️ Usage

Command Line

python main.py \
  --image <base_image_path> \
  --prompt "Your descriptive prompt here" \
  --n <iterations> \
  [--retries <X>]

• --image (-i): path to your starting image file (PNG/JPG/WebP).
• --prompt (-p): text prompt guiding the edits.
• --n (-n): number of iterations (chain length).
• --retries (-r, optional): automatic retry count on failure. If omitted, script will prompt you interactively.

Gradio Web App

You can use a web interface for uploading images and entering prompts:

python app.py

• The app will be available at http://localhost:7860 by default.
• To make it accessible on your network, it binds to 0.0.0.0:7860.
• Make sure your OPENAI_API_KEY is set in your environment.
• Features include:
  • Image upload and prompt entry
  • Iteration and retry controls
  • Animated GIF of the generation process
  • Gallery view of all intermediate steps
  • Final result display

Docker

Build and run the app in a container:

docker build --build-arg OPENAI_API_KEY=your-api-key-here -t modelcollapse .
docker run -p 7860:7860 modelcollapse

• The app will be available at http://localhost:7860
• The API key is passed securely as a build argument and set as an environment variable in the container.

---

🔄 Customization Points

In main.py, you can tweak these parameters directly:

Parameter	Default	Description
QUALITY	high	Image quality level (low, medium, high or auto).
SIZE	1024x1024	Output resolution. Options: 1024x1024, 1536x1024 (landscape), 1024x1536 (portrait), auto
RETRIES	None	Set default retry count; passing --retries overrides this.

To change, locate the client.images.edit(...) call and modify the parameters accordingly.

---

⚠️ Notes

• The script expects your OPENAI_API_KEY in the environment. No hard-coded keys.
• If an iteration fails and retries are exhausted (or you choose not to retry), the chain stops early and returns whatever was generated up to that point.
• The Gradio app allows for easy testing and demoing in a browser.
• For Docker, always pass your API key securely as shown above.

---

📄 License

MIT