[Paper Review] Editing Implicit Assumptions in Text-to-Image Diffusion Models

Hadas Orgad, Bahjat Kawar+1 more|arXiv (Cornell University)|Mar 14, 2023

Generative Adversarial Networks and Image Synthesis4 citations

TL;DR

This paper proposes TIME (Text-to-Image Model Editing), a highly efficient method to edit implicit world assumptions in text-to-image diffusion models by modifying only 2.2% of the model's parameters. It edits cross-attention projection matrices to align a source prompt (e.g., 'a pack of roses') with a destination prompt (e.g., 'a pack of blue roses'), enabling rapid, non-forgetting edits in under a second on a single GPU while preserving generation quality and generalization to related prompts.

ABSTRACT

Text-to-image diffusion models often make implicit assumptions about the world when generating images. While some assumptions are useful (e.g., the sky is blue), they can also be outdated, incorrect, or reflective of social biases present in the training data. Thus, there is a need to control these assumptions without requiring explicit user input or costly re-training. In this work, we aim to edit a given implicit assumption in a pre-trained diffusion model. Our Text-to-Image Model Editing method, TIME for short, receives a pair of inputs: a "source" under-specified prompt for which the model makes an implicit assumption (e.g., "a pack of roses"), and a "destination" prompt that describes the same setting, but with a specified desired attribute (e.g., "a pack of blue roses"). TIME then updates the model's cross-attention layers, as these layers assign visual meaning to textual tokens. We edit the projection matrices in these layers such that the source prompt is projected close to the destination prompt. Our method is highly efficient, as it modifies a mere 2.2% of the model's parameters in under one second. To evaluate model editing approaches, we introduce TIMED (TIME Dataset), containing 147 source and destination prompt pairs from various domains. Our experiments (using Stable Diffusion) show that TIME is successful in model editing, generalizes well for related prompts unseen during editing, and imposes minimal effect on unrelated generations.

Motivation & Objective

• To address the problem of text-to-image diffusion models learning incorrect, outdated, or biased implicit assumptions from training data.
• To enable editing of these assumptions without retraining, fine-tuning, or explicit user input per generation.
• To develop a method that modifies only a small fraction of model parameters while preserving overall generative quality and avoiding catastrophic forgetting.
• To create a benchmark dataset (TIMED) for evaluating model editing performance across diverse domains and concepts.
• To ensure edits generalize to related prompts while leaving unrelated generations unaffected.

Proposed method

• TIME edits the projection matrices in cross-attention layers of the diffusion model, which map textual tokens to visual features.
• It computes a parameter update that minimizes the distance between the embedding of the source prompt and the destination prompt in the cross-attention space.
• The method uses a single optimization step with a learned hyperparameter λp to balance the edit strength and original weight fidelity.
• The edit is applied in parallel across all cross-attention layers and requires no fine-tuning or data re-collection.
• The text encoder remains unmodified, ensuring the language understanding capacity is preserved.
• The approach is computationally efficient, modifying only 2.2% of the diffusion model’s parameters and completing in under one second on a consumer GPU.

Experimental results

Research questions

• RQ1 Can implicit assumptions in text-to-image diffusion models be edited efficiently without retraining or fine-tuning?
• RQ2 How well does the edit generalize to related prompts not seen during editing?
• RQ3 Does the editing process preserve the model’s overall generative quality and performance on unrelated prompts?
• RQ4 To what extent can the method mitigate social biases, such as gender stereotypes in profession representations?
• RQ5 How effective is the method in changing model behavior for specific attributes (e.g., color of roses) while maintaining fidelity?

Key findings

• TIME successfully edits implicit assumptions in Stable Diffusion, reducing the percentage of female figures in stereotypically male professions from a baseline of 1.6% to 11.2% on average, with a Δp of 0.15 or lower for most professions.
• For the profession 'nurse', the baseline had 100% female representation, which was reduced to 92.0% after editing, with Δp = 0.83 ± 0.05.
• The method generalizes well: after editing 'roses' to be blue, related prompts like 'a field of roses' now generate blue roses, while unrelated prompts like 'a poppy field' remain unchanged.
• The edit is highly efficient, modifying only 2.2% of the model’s parameters and completing in under one second on a single consumer-grade GPU.
• FID scores remain stable post-editing, indicating no degradation in image quality or generative capability.
• The method achieves significant bias mitigation across 147 prompt pairs in the TIMED dataset, with minimal impact on unrelated generations.