Open Source AI Coalition Plans New Model to Compete With Stable Diffusion

A newly fashioned alliance of open-source AI builders has bold plans to launch a brand new AI mannequin, one that might compete immediately with Secure Diffusion and its controversial newest launch.

The Open Mannequin Initiative (OMI) was launched final week by Invoke, a Generative AI platform for Skilled Studios, Cozy.org, which focuses on the event of ComfyUI, and CivitAI, the world’s largest Secure Diffusion repository. Now, the group is spearheading a community-driven effort aimed toward growing open-source AI fashions for picture, video, and audio technology.

“I hope to have the OMI launch a brand new, next-gen mannequin that the group can rally behind and construct help for within the subsequent three to 6 months—an bold purpose for a brand new group, however one which I’m assured we will obtain,” Kent Keirsey, CEO of Invoke AI, instructed Decrypt.

The OMI goals to develop open-source fashions of equal or higher high quality than proprietary fashions (like Ideogram or MidJourney) however freed from restrictive licensing phrases.

“Cozy is inclined to leverage a better-designed SD3.” CivitAI declared in a current submit.

This initiative comes shortly after the launch of SD3 by Stability AI, which was criticized for its restrictive licensing phrases. Whereas the Secure Diffusion household of AI fashions has been massively in style, the SD3 license was deemed so restrictive that it led to its ban from CivitAI.

To forestall this from occurring once more, the group advocates for actually open-source growth and can possible undertake the MIT or Apache-2 license, which ensures actually open, non-restrictive fashions topic to minimal situations.

“We imagine open supply is one of the simplest ways ahead to make sure that AI advantages everybody,” the members mentioned in an open letter. “By teaming up, we will ship high-quality, aggressive fashions with open licenses that push AI creativity ahead, are free to make use of, and meet the wants of the group.”

Addressing moral considerations, the initiative additionally dedicated to growing a base mannequin with out pre-trained capabilities corresponding to “recognition of unconsented artist names” and “producing the likeness of unconsented people.”

The initiative has garnered important help, with over 1,000 members becoming a member of its Discord server. “We’ve fielded greater than 100 requests to affix and help in simply the final 24 hours,” Keirsey instructed Decrypt.

Concerning funding, Keirsey mentioned that the initiative wouldn’t pursue enterprise capital however quite depend on group help and the enterprise fashions of its founding members.

“The OMI won’t be taking investments, because the initiative’s goal is constructing open fashions, not producing a revenue,” he mentioned, “We’ve seen what occurs when revenue turns into the motivating issue behind organizations based on the precept of open entry to AI.”

Every member of the OMI will preserve its personal enterprise construction regardless of its involvement with the group.

Robin Ken of ComfyUI instructed Decrypt that their function within the mission of “democratizing AI” is on the tooling layer, not like different initiatives like Stability, Mistral or Meta, that are extra centered on growing fashions. He additionally confirmed that Cozy is greater than a aspect undertaking and has monetary backing behind it.

“Cozy will stay open supply endlessly,” Ken instructed Decrypt. “We’re VC-backed and plan to make cash from consulting/enterprise help.”

Ken, Alex Goodwin, and pseudonymous cofounder Cozy Nameless left Stability AI to give attention to the event and progress of Cozy.org, together with different builders.

The Open Mannequin Initiative mentioned it’s going to initially give attention to organizing volunteers, deciding on a governance construction, and curating datasets with group help.

LAION—an AI agency that compiled datasets of pictures and captions scraped from the online that had been used to coach a number of Secure Diffusion—was initially introduced as a founding member of the OMI however was later eliminated.

Edited by Ryan Ozawa.