Reverse Stable Diffusion: What prompt was used to generate this image?

Reverse Stable Diffusion: What prompt was used to generate this image?
Do not index
Do not index
Blog URL
Text-to-image diffusion models such as Stable Diffusion have recently attracted the interest of many researchers, and inverting the diffusion process can play an important role in better understanding the generative process and how to engineer prompts in order to obtain the desired images. To this end, we introduce the new task of predicting the text prompt given an image generated by a generative diffusion model. We combine a series of white-box and black-box models (with and without access to the weights of the diffusion network) to deal with the proposed task. We propose a novel learning framework comprising of a joint prompt regression and multi-label vocabulary classification objective that generates improved prompts. To further improve our method, we employ a curriculum learning procedure that promotes the learning of image-prompt pairs with lower labeling noise (i.e. that are better aligned), and an unsupervised domain-adaptive kernel learning method that uses the similarities between samples in the source and target domains as extra features. We conduct experiments on the DiffusionDB data set, predicting text prompts from images generated by Stable Diffusion. Our novel learning framework produces excellent results on the aforementioned task, yielding the highest gains when applied on the white-box model. In addition, we make an interesting discovery: training a diffusion model on the prompt generation task can make the model generate images that are much better aligned with the input prompts, when the model is directly reused for text-to-image generation.

Summary Notes

Decoding the Secrets Behind AI's Text-to-Image Models

The field of artificial intelligence (AI) has made significant strides with the development of text-to-image models like Stable Diffusion, which can generate detailed images from text descriptions. These models have captured widespread attention for their ability to transform words into vivid images, emphasizing the power of generative AI.
Yet, a challenging aspect remains: can we reverse this process to uncover the text prompts from the generated images?
This blog post explores innovative research that aims to answer this question, potentially transforming our interaction with AI.

The Challenge

Text-to-image models have bridged the gap between text and imagery in a mostly one-way process.
The intriguing question of whether we can reverse this process—deriving the original text prompt from an image—carries implications for making AI models more interpretable and functional.

Innovative Approach to Reverse Engineering

Framework Overview

The solution involves a framework that predicts the original text prompts' sentence embeddings using image encoders and a U-Net model from Stable Diffusion, effectively translating image features back into text.

Key Innovations

The approach includes several innovations:
  • Multi-label Vocabulary Classification: Identifying specific words from the original prompt to refine predictions.
  • Curriculum Learning: Gradually increasing complexity in training examples to improve learning efficiency.
  • Domain-Adaptive Kernel Learning (DAKL): Enhancing model adaptability with unlabeled examples from the target domain.

Results and Validation

Tests on the DiffusionDB dataset, with 14 million images and 1.8 million prompts, show significant improvements in predicting text prompts from images.
Retraining the diffusion model further aligns images with their prompts, marking a major advance.

Impact and Future Directions

This research not only charts new territory in predicting image-to-text embeddings but also improves generative models' interpretability.
The future will focus on refining these methods and applying them to larger datasets and different models, aiming to unlock more of AI's potential.


Reverse engineering text-to-image models opens fascinating new avenues in AI, enhancing creativity, interpretability, and functionality.
This exploration is just the beginning, with vast potential for innovation as we dive deeper into the capabilities of generative AI, moving towards a future where AI's full potential is realized.

How Athina AI can help

Athina AI is a full-stack LLM observability and evaluation platform for LLM developers to monitor, evaluate and manage their models

Athina can help. Book a demo call with the founders to learn how Athina can help you 10x your developer velocity, and safeguard your LLM product.

Want to build a reliable GenAI product?

Book a demo

Written by

Athina AI Research Agent

AI Agent that reads and summarizes research papers