Precision in Pixels: NVIDIA’s Edify Image Model Combines High Quality with Unmatched Control

The field of text-to-image synthesis has advanced rapidly, with state-of-the-art models now generating highly realistic and diverse images from text descriptions. This progress largely owes to diffusion-based architectures trained on vast datasets containing billions of image-text pairs. The capability to create high-resolution, photorealistic images from text has transformative potential across fields like content creation, gaming, synthetic data generation, and digital avatar design.

In a new paper Edify Image: High-Quality Image Generation with Pixel Space Laplacian Diffusion Models, an NVIDIA research team introduces Edify Image—a suite of pixel-based diffusion models that achieve high-resolution image synthesis with exceptional control and precision.

Unlike traditional pixel-space generators, which stack diffusion stages to progressively upscale low-resolution images, often resulting in artifacts, Edify Image synthesizes large image contexts through a single, cohesive diffusion process.

The key innovation of Edify Image is the Laplacian Diffusion Model, a multi-scale diffusion approach that decays image frequency bands at varying rates over time. This approach enables the model to manage resolution in a way that minimizes artifacts, preserving high-quality detail across the entire image. The researchers implemented this model within a U-Net-based architecture, using residual and attention blocks that sequentially downsample and upsample feature maps with skip connections.

For added efficiency in high-resolution generation, they introduced invertible wavelet transforms at the beginning and end of the network, using 2-level Haar wavelets to downsample images in pixel space. This optimization reduces the number of spatial tokens in the attention layers by a factor of 16, substantially enhancing computational efficiency during training.

Empirical results demonstrate Edify Image’s ability to generate images with flexible aspect ratios based on detailed text prompts, along with improvements in fairness, diversity, and the capacity to incorporate camera controls like pitch and depth of field. Edify Image can produce high-frequency details that remain true to low-resolution inputs, generating top-quality images while supporting flexible structural adjustments.

Edify Image’s innovative approach to multi-scale diffusion and efficient use of pixel space represent a significant step forward in text-to-image synthesis, enabling controllable, high-quality image generation with wide-ranging applications across creative and technical domains.

The paper Edify Image: High-Quality Image Generation with Pixel Space Laplacian Diffusion Models is on arXiv.

---

Author: Hecate He | Editor: Chain Zhang

---