Granted, the biggest blocker to making money off of VQGAN + CLIP in a scalable manner is generation speed unlike most commercial AI models which use inference and can therefore be optimized to drastically increase performance, VQGAN + CLIP requires training, which is much slower and can’t allow content generation in real time like GPT-3. There are other VQGANs available such as ones trained on the Open Images Dataset or COCO, both of which have commercial-friendly CC-BY-4.0 licenses, although in my testing they had substantially lower image generation quality. However, the ImageNet 16384 VQGAN used in this Colab Notebook and many other VQGAN+CLIP Notebooks was trained on ImageNet, which has famously complicated licensing, and whether finetuning the VQGAN counts as sufficiently detached from an IP perspective hasn’t been legally tested to my knowledge.
#Clip art for mac keynote license
In contrast to StyleGAN2 images (where the license is explicitly noncommercial), all aspects of the VQGAN + CLIP pipeline are MIT Licensed which does support commericalization.
Making Money Off of VQGAN + CLIPĬan these AI generated images be commercialized as software-as-a-service? It’s unclear. Reality is an illusion:8 - fas fa-eye icon, icon initial image, icon target image, black icon background, white icon, learning rate = released an album with many more examples of prompt augmentations and their results. Some say art is better when there’s mystery, but my view is that knowing how AI art is made is the key to making even better AI art. The VQGAN + CLIP technology is now in a good state such that it can be used for more serious experimentation. created a fork of that Notebook which has a user-friendly UI, to which I became aware of how far AI image generation has developed in a few months.įrom that, I forked my own Colab Notebook, and streamlined the UI a bit to minimize the number of clicks needs to start generating and make it more mobile-friendly.
Twitter accounts like and which leverage VQGAN + CLIP with user-submitted prompts have gone viral and received mainstream press. The core CLIP-guided training was improved and translated to a Colab Notebook by Katherine Crawson/ and others in a special Discord server.
Then open source worked its magic: the GAN base was changed to VQGAN, a newer model architecture Patrick Esser and Robin Rombach and Björn Ommer which allows more coherent image generation. The first implementation was Big Sleep by Ryan Murdock/ which combined CLIP with an image generating GAN named BigGAN. Since CLIP is essentially an interface between representations of text and image data, clever hacking can allow anyone to create their own pseudo-DALL-E.
0 kommentar(er)
