AI generates better type and graphics - DeepFloyd/IF

[15:02 Tue,2.May 2023 by Rudi Schmidts]

The company Stability AI (which among other things also significantly promotes the open source Stable-Diffusion) has introduced another image generator with DeepFloyd/IF. It is said to be particularly suitable for fonts and graphics.

Those who have already gained experience with diffusion-based AI image generators know the problem: Correct type - no matter in which language - is practically impossible to generate. What you see is usually - if at all - a linguistic gibberish of hallucinated letters.

Stable Diffusions language seems to be out of this world.

But this problem shall be over now, because the new DeepFloyd/IF model shall allow photorealistic representations with lettering. It is also said to be particularly suitable for graphic tasks such as logo design.

DeepFloyd is based on Google&s AI image generator Imagen. This works somewhat differently than Stable Diffusion and combines an open source large language model (LLM) from Google ( T5-XXL-1.1) with a pixel diffusion model.

The latter works in three stages and primarily generates only 64 x 64 pixel images, which are then scaled up twice via superresolution over 256 x 256 pixels to the output resolution of 1024 x 1024 pixels. The image generator was trained with the proven LAION-A data set with 1.2 billion images.

DeepFloyd/IF can generate readable text and graphics

An official web image generator to try out DeepFloyd/IF online does not exist yet - because the current license only allows the use for research and not for commercial purposes. However, if you want to "research" it yourself, you can find

on Github corresponding packages for download.

At the same time, however, DeepFloyd/IF also heralds a new era for AI home use. Because while previous Stable Diffusion models already work with graphics cards from about 6 GB memory, DeepFloyd now requires at least 16 GB GPU memory. For the higher-quality (and thus larger) model, even 24 GB are mandatory. Such strongly increasing requirements for GPU memory in the upcoming AI applications

we had already discussed this at slashCAM recently.

more infos at bei github.com

deutsche Version dieser Seite: Schluss mit Kauderwelsch - neue Bild-KI DeepFloyd / IF kann auch schreiben