The list of generative AI image models just keeps growing.
On Monday, AI company Luma released Photon, its new text-to-image model. Geared towards designers and creatives, the model builds off of Luma’s Dream Machine video generator, which lets users create short video clips from images they upload.
Also: The best AI image generators of 2024
Like Dream Machine, Photon is designed for everyone and targeted towards those who “build the world around us, from designers, movie makers, visual thinkers to small businesses and educators,” the release explains. This includes use cases for everything from fashion to architecture.
Featuring multi-image output and sharp text rendering, Photon is built on Luma’s bespoke Universal Transformer architecture, which the company claims makes it “800% faster and cheaper than other comparable models.” Like similar generators, users can tweak their prompts as they go and edit and customize the resulting images.
Luma also announced a Photon API, that may be released before the end of 2024. The scalable API will let developers create text-to-image and image-to-image content.
“The real magic of Luma AI is in becoming the home for each creator’s unique creative DNA — their references, their style, their history — while Dream Machine learns and evolves with each user’s creative journey,” Anjney Midha, General Partner at a16z and Luma AI board member, said in the release.
Also: Two free ways to get a Perplexity Pro subscription for one year
Luma wants users to think of Dream Machine and Photon as true creative collaborators that can understand users’ unique contexts and needs. One feature allows users to upload a photo of a person and create a reusable “character” of them that Photon can depict across several single-image scenarios — an ideal tool for brainstorming campaigns. Users can upload the person’s likeness and include “@character” in their prompt to activate the feature.
However, during a demo for ZDNET, Photon appeared to have some issues rendering natural-looking expressions for characters in certain positions, which immediately brings to mind concerns about deepfakes and other potential misuses.
Luma’s usage guidelines include a note on content moderation that indicates the company will remove certain generated images.
Also: AI isn’t hitting a wall, it’s just getting too smart for benchmarks, says Anthropic
Notably, Luma states that input and output data from Photon aren’t used to train the company’s models. When asked about Luma’s training dataset during the demo, CEO Amit Jain said he couldn’t divulge what it includes, and whether it contains material scraped from the internet or licensed content.
In terms of copyright concerns, the guidelines are a bit more vague. The company asks that users “do not request or generate content that infringes on copyrighted or trademarked materials” and “always respect intellectual property rights.”
Lastly, the company also launched a new interface for Dream Machine, which is available now on the web and iOS. According to the company, it can be used with natural language: no need to learn how to be a better prompt engineer.
Also: This AI image generator that went viral for its realistic images gets a major upgrade
The new interface is easy to use and reminiscent of other generators that identify key descriptors in your prompt and offer a drop-down menu of similar alternatives in case you’d like to modify your idea or explore different variations. The service is geared towards any user who wants to explore a design concept, iterate on a product idea, brainstorm, and collaborate with team members via Luma’s handy mood board infrastructure.
The “visual AI service” also offers personalization features for images and video, as well as free and paid subscription tiers starting at $10 a month.
Artificial Intelligence
