This AI Research Unveils ‘Kandinsky1’: A New Approach in Latent Diffusion Text-to-Image Generation with Outstanding FID Scores on COCO-30K

In recent years, computer vision and generative modeling have witnessed remarkable progress, leading to advancements in text-to-image generation. Various generative architectures, including diffusion-based models, have played a pivotal role in enhancing the quality and diversity of generated images. This article explores the principles, features, and capabilities of Kandinsky1, a powerful model with 3.3 billion parameters, and highlights its top-tier performance in measurable image generation quality.

Text-to-image generative models have evolved from autoregressive approaches with content-level artifacts to diffusion-based models like DALL-E 2 and Imagen. These diffusion models, categorized as pixel-level and latent-level, excel in image generation, surpassing GANs in fidelity and diversity. They integrate text conditions without adversarial training, demonstrated by models like GLIDE and eDiff-I, which generate low-resolution images and upscale them using super-resolution diffusion models. These advancements have transformed text-to-image generation.

Researchers from AIRI, Skoltech, and Sber AI introduce Kandinsky, introduce a novel text-to-image generative model (Kandinsky) that combines latent diffusion techniques with image prior models. Kandinsky features a modified MoVQ implementation as its image autoencoder component and separately trains the image prior model to map text embeddings to CLIP’s image embeddings. Their method provides a user-friendly demo system supporting diverse generative modes and releases the model’s source code and checkpoints. 

Their approach introduces a latent diffusion architecture for text-to-image synthesis, leveraging image prior models and latent diffusion techniques. It employs an image-prior approach that incorporates diffusion and linear mappings between text and image embeddings using CLIP and XLMR text embeddings. Their model comprises three key steps: text encoding, embedding mapping (image prior), and latent diffusion. Elementwise normalization of visual embeddings based on full-dataset statistics is implemented, expediting the convergence of the diffusion process.

The Kandinsky architecture performs strongly in text-to-image generation, attaining an impressive FID score of 8.03 on the COCO-30K validation dataset at a resolution of 256×256. The Linear Prior configuration yielded the best FID score, indicating a potential linear relationship between visual and textual embeddings. Their model’s proficiency is demonstrated by training a “cat prior” on a subset of cat images, which excelled in image generation. Overall, Kandinsky competes closely with state-of-the-art models in text-to-image synthesis.

Kandinsky, a latent diffusion-based system, emerges as a state-of-the-art performer in image generation and processing tasks. Their research extensively explores image prior design choices, with the linear prior showing promise and hinting at a linear connection between visual and textual embeddings. User-friendly interfaces like a web app and Telegram bot facilitate accessibility. Future research avenues encompass leveraging advanced image encoders, enhancing UNet architectures, improving text prompts, generating higher-resolution images, and exploring features like local editing and physics-based control. Researchers underscore the need to address content concerns, suggesting real-time moderation or robust classifiers for mitigating undesirable outputs.

Check out the Paper and Github. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 31k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..

We are also on WhatsApp. Join our AI Channel on Whatsapp..