metadata

pipeline_tag: text-to-image
widget:
  - text: >-
      high quality, portrait photo of 30 y.o european man, wearing black shirt,
      serious face, cinematic shot, dramatic lighting 
    output:
      url: 1man.png
  - text: 'A tiny robot taking a break under a tree in the garden '
    output:
      url: robot.png
  - text: mystery
    output:
      url: mystery.png
  - text: a cat wearing sunglasses in the summer
    output:
      url: cat.png
  - text: 'robot holding a sign that says ’a storm is coming’ '
    output:
      url: storm.png
  - text: the vibrance of the human soul
    output:
      url: soul.png
  - text: Lady of War, chique dark clothes, vinyl, imposing pose, anime style, 90s
    output:
      url: anime.png
license: cc-by-nc-nd-4.0

Constructive Deconstruction: Domain-Agnostic Debiasing of Diffusion Models

Introduction

Constructive Deconstruction is a novel approach to debiasing diffusion models used in generative tasks like image synthesis. This method enhances the quality and fidelity of generated images across various domains by removing biases inherited from the training data. Our technique involves overtraining the model to a controlled noisy state, applying nightshading, and using bucketing techniques to realign the model's internal representations.

Methodology

Overtraining to Controlled Noisy State

By purposely overtraining the model until it predictably fails, we create a controlled noisy state. This state helps in identifying and addressing the inherent biases in the model's training data.

Nightshading

Nightshading is repurposed to induce a controlled failure, making it easier to retrain the model. This involves injecting carefully selected data points to stress the model and cause predictable failures.

Bucketing

Using mathematical techniques like slerp (Spherical Linear Interpolation) and bislerp (Bilinear Interpolation), we merge the induced noise back into the model. This step highlights the model's learned knowledge while suppressing biases.

Retraining and Fine-Tuning

The noisy state is retrained on a large, diverse dataset to create a new base model called "Mobius." Initial issues such as grainy details and inconsistent colors are resolved during fine-tuning, resulting in high-quality, unbiased outputs.

Results and Highlights

Increased Diversity of Outputs

Training the model on high-quality data naturally increases the diversity of the generated outputs without intentionally loosening associations. This leads to improved generalization and variety in generated images.

Empirical Validation

Extensive experiments and fine-tuning demonstrate the effectiveness of our method, resulting in high-quality, unbiased outputs across various styles and domains.