Google Launches Muse, A New Text-to-Image Transformer Model

Muse claims to be faster as it uses a compressed, discrete latent space and parallel decoding.
Listen to this story

Since the beginning of 2021, advances in AI research have been revolutionised with the birth of a plethora of deep learning-backed text-to-image models like DALL-E-2, Stable Diffusion, and Midjourney, to name a few. Adding to the list is Google’s Muse, a text-to-image Transformer model that claims to achieve state-of-the-art image generation performance. 

Given the text embedding obtained from a large language model (LLM) that has already been trained, Muse is trained on a masked modelling task in discrete token space. Muse has been trained to predict randomly masked image tokens. Muse asserts to be more effective than pixel-space diffusion models like Imagen and DALL-E 2 since it uses discrete tokens and requires fewer sample iterations. The model generates a zero-shot, mask-free editing for free by iteratively resampling image tokens conditioned on a text prompt.

Check out more about Muse here

AIM Daily XO

Join our editors every weekday evening as they steer you through the most significant news of the day, introduce you to fresh perspectives, and provide unexpected moments of joy
Your newsletter subscriptions are subject to AIM Privacy Policy and Terms and Conditions.

Unlike Parti and other autoregressive models, Muse uses parallel decoding. A pre-trained LLM enables fine-grained language understanding, translating to high-fidelity image generation and comprehending visual concepts such as objects, their spatial relationships, pose, cardinality, etc. Additionally, Muse supports inpainting, outpainting, and mask-free editing without the need to modify or invert the model.

With an FID score of 6.06, the 900M parameter model achieves a new SOTA on CC3M. On zero-shot COCO evaluation, the Muse 3B parameter model obtains an FID of 7.88 and a CLIP score of 0.32. 

Download our Mobile App

Model Architecture:

For both the base and super-res Transformer layers, the text encoder creates a text embedding that is used for cross-attention with image tokens. The base model then uses a VQ Tokenizer that generates a 16*16 latent space of tokens after being pre-trained on lower resolution (256*256) images. The cross-entropy loss then learns to predict the masked tokens that have been masked at a variable rate for each sample. After training the base model, the reconstructed lower-resolution tokens and text tokens are then fed into the super-res model. Now the model can predict masked tokens at a higher resolution.

Sign up for The Deep Learning Podcast

by Vijayalakshmi Anandan

The Deep Learning Curve is a technology-based podcast hosted by Vijayalakshmi Anandan - Video Presenter and Podcaster at Analytics India Magazine. This podcast is the narrator's journey of curiosity and discovery in the world of technology.

Shritama Saha
Shritama is a technology journalist who is keen to learn about AI and analytics play. A graduate in mass communication, she is passionate to explore the influence of data science on fashion, drug development, films, and art.

Our Upcoming Events

24th Mar, 2023 | Webinar
Women-in-Tech: Are you ready for the Techade

27-28th Apr, 2023 I Bangalore
Data Engineering Summit (DES) 2023

23 Jun, 2023 | Bangalore
MachineCon India 2023 [AI100 Awards]

21 Jul, 2023 | New York
MachineCon USA 2023 [AI100 Awards]

3 Ways to Join our Community

Telegram group

Discover special offers, top stories, upcoming events, and more.

Discord Server

Stay Connected with a larger ecosystem of data science and ML Professionals

Subscribe to our Daily newsletter

Get our daily awesome stories & videos in your inbox

Council Post: The Rise of Generative AI and Living Content

In this era of content, the use of technology, such as AI and data analytics, is becoming increasingly important as it can help content creators personalise their content, improve its quality, and reach their target audience with greater efficacy. AI writing has arrived and is here to stay. Once we overcome the initial need to cling to our conventional methods, we can begin to be more receptive to the tremendous opportunities that these technologies present.