Glass wants to enhance the software and technology of your phone’s camera. Despite its wacky anamorphic lens, the business (with $9.3 million in fresh funding) has produced an AI-powered camera update that it claims greatly increases picture quality without AI upscaling errors.
GlassAI’s neural image signal processor improves photos using software. ISPs convert flat, noisy, and distorted sensor inputs into clear, vivid pictures.
The ISP is becoming more complicated, as phone companies like Apple and Google demonstrate, synthesizing several exposures, fast identifying and sharpening faces, correcting for minor motions, etc. Many use machine learning or AI, but they must be careful: AI-generated detail may cause hallucinations or artifacts by trying to build visual information from nothing. Some “super-resolution” models are valuable, but they must be controlled.
Glass produces a whole camera system with a lozenge-shaped front element and an ISP. While the former is preparing for the market with new gadgets, the latter is a viable product.
“Our restoration networks correct optical aberrations and sensor issues while efficiently removing noise and outperform traditional Image Signal Processing pipelines at fine texture recovery,” said CTO and co-founder Tom Bishop in their press statement.
The phrase “recovery” is important since details are taken from raw images. Depending on your camera stack, you may be able to rectify or exploit artifacts, angles, or noise patterns. Computational photography stacks include learning how to transform suggested features into actual ones or combine information from many exposures. CEO and co-founder Ziv Attar claims their neural ISP is industry-leading.
He noted that Apple only uses a neural image stack when necessary and gets poor results. He showed how Glass outperformed Apple’s neural ISP in text interpretation:
“I think it’s fair to assume that if Apple hasn’t managed to get decent results, it is a hard problem to solve,” he added. More important than the stack is how you train. Our unique method, created for anamorphic lens systems, works well on any camera. Our training laboratories use robotics and optical calibration equipment to teach a network to assess lens aberration in a comprehensive manner and reverse optical distortion.
He cited a Moto Edge 40 camera test by DXO and GlassAI. All glass-processed photographs are enhanced, sometimes greatly.
In night mode, the built-in ISP fails to distinguish tiny lines, textures, and facial characteristics. Even with half the exposure time, GlassAI makes it crisp.
Switching between raws and finals lets you see pixels in several test photographs Glass has.
Phone and camera manufacturers must spend a lot of effort tweaking the ISP to ensure the sensor, lens, and other parts work together to provide the best picture. However, Glass’s one-size-fits-all approach may work better in less time.
“Training shippable software on new devices can take anywhere from hours to days. Phone companies spend months refining image quality with large teams. “Our fully automated process supports multiple devices in a few days,” added Attar.
The neural ISP is end-to-end, going from sensor raw to finished picture without denoising, sharpening, etc.
Attar clarified that their approach is distinct from super-resolution AI services, which upscale final images. These aren’t always “recovering” information, but creating them when appropriate, which might have unintended consequences. Glass incorporates AI, although it’s not generative like image-related AIs.
The product is available to the public today, likely after extensive partner testing. It may be worth trying if you create an Android phone.
The phone with the lozenge-shaped anamorphic camera won’t be released until that manufacturer is ready.
Glass has been raising funds while developing its technology and testing clients. The firm secured a $9.3 million “extended seed,” since the seed round was in 2021. GV led the fundraising, alongside Future Ventures, Abstract Ventures, and LDV Capital.
