Google's Parti: The Newest Text-To-Image Generation Model
Google has revealed the newest in text-to-image generation models today with Parti.
Created on June 22|Last edited on June 23
Comment
Text-to-image generation models are all the rage these days, but even this announcement is sudden and a little unexpected. Following so closely the announcement of Imagen, Google research's previous image-to-text generation model, they're deciding to show off another model build to do the same thing.
This newest model in the spotlight today is named Parti (Pathways Autoregressive Text-to-Image). While Imagen and DALL·E 2 are diffusion models, Parti follows in DALL·E's footsteps as an autoregressive model. Regardless of architecture and training method, the end use is the same: These models, including Parti, will generate detailed images based on the user's text input.
Accompanying the revealing of Parti is a blog post describing the process of creating images with Google's text-to-image models, available here: https://blog.google/technology/research/how-ai-creates-photorealistic-images-from-text/
The Parti details
Four model sizes of Parti were created by the researchers, including parameter counts at 350 million, 750 million, 3 billion, and 20 billion. These models were trained using Google Cloud TPUs which were able to easily support the creation of these huge model sizes. Several comparisons between the model sizes are provided on the website, but ill share a couple from the paper here (smallest to largest from left to right):
Like all the other text-to-image generators out there, Parti struggles in a variety of similar ways. Incorrect object counts, blended features, incorrect relational positioning or size, not handling negation correctly, the list could go on. Here are a few examples of what Parti struggles at:
Can I join the Parti?
So, you're probably wondering if maybe this time we'll get to play with this month's newest big text-to-image generator? Unsurprisingly, the answer is no, and we don't even get some hoops to jump through for access in the form of a waitlist you can sign up for.
All the regular reasons apply: Bias in training data, fear of creating harmful imagery, inevitable misuse by the public, etc.
Find out more
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.