Home AI News The Power of FABRIC: Integrating Feedback into Generative AI for Image Synthesis

The Power of FABRIC: Integrating Feedback into Generative AI for Image Synthesis

The Power of FABRIC: Integrating Feedback into Generative AI for Image Synthesis

Generative AI: Understanding Diffusion Models and the Role of FABRIC

Generative AI has become a well-known term in recent years, playing a crucial role in various applications. One of the most powerful types of generative models is diffusion models. These models have revolutionized image synthesis and related tasks by generating high-quality and diverse images. Unlike traditional models like GANs and VAEs, diffusion models refine a noise source to create stable and coherent images.

The Significance of Diffusion Models

Diffusion models have gained significant popularity due to their ability to generate high-fidelity images while maintaining stability during training. They have been widely adopted for image synthesis, inpainting, and style transfer tasks.

The Challenge with Diffusion Models

Despite their impressive capabilities, diffusion models face challenges when it comes to generating specific desired outputs based on textual descriptions. It can be frustrating to accurately describe preferences through text prompts, as the model may ignore them or not fully understand. This often requires users to refine the generated image to make it usable.

Introducing FABRIC: Feedback via Attention-Based Image Conditioning

FABRIC (Feedback via Attention-Based Reference Image Conditioning) is a novel approach that addresses this challenge by allowing users to provide iterative feedback during the generative process of diffusion models.

How FABRIC Works

FABRIC leverages positive and negative feedback images gathered from previous generations or human input. It utilizes reference image-conditioning to refine future results, creating a more controllable and interactive text-to-image generation process. This approach is inspired by ControlNet, which enables the generation of new images similar to reference images. FABRIC incorporates a self-attention module in the U-Net architecture, allowing the model to “pay attention” to other pixels in the image and inject additional information from a reference image. The keys and values for reference injection are computed through the U-Net of Stable Diffusion, enabling the denoising process to incorporate semantic information from the reference image.

Iterative Refinement with FABRIC

FABRIC is further extended to incorporate multi-round positive and negative feedback. Separate U-Net passes are performed for each liked and disliked image, and the attention scores are reweighted based on the feedback received. The feedback process can be scheduled according to denoising steps, allowing for the iterative refinement of generated images.


FABRIC provides a solution to the challenge of steering diffusion models towards specific desired outputs. By integrating iterative feedback into the generative process, users can have more control over the image generation pipeline. FABRIC’s ability to leverage reference image-conditioning and incorporate user preferences enhances the quality and usability of generated images. This approach opens up new possibilities for interactive text-to-image generation in AI.

Source link


Please enter your comment!
Please enter your name here