[10:33 Sun,19.January 2025 by Thomas Richter] |
An image in under a second? Nvidia&s new image AI Sana can do just that: on a laptop GPU with 16 GB VRAM, a 1,024 x 1,024 resolution image takes less than a second to generate. Pleasingly, Nvidia has also released Sana&s source code under the free Apache 2.0 license, meaning users – provided they have a powerful graphics card – can use Sana&s speed on their own computer and generate images virtually for free. ;)
Nvidia compared Sana&s speed to  FLUX.1 (Flux), currently the best and most popular free image AI that can also be run locally. According to the benchmarks, Sana (version 0.6B) is 20x smaller and about 100x faster at generating images in 4K resolution (4,096 x 4,096) or 25x faster at 1,024 x 1,024 images than the FLUX.1-12B model, which is open for non-commercial use and offers higher quality than the smaller and faster FLUX.1. As with other image AIs, the 4K image is a KI-upscaled version (in this case SUPIR) of an image originally generated at 1,024 x 1,024. The Sana team has also developed a plugin for integration with ComfyUI. ;) Speed Unfortunately, Nvidia has not published an ELO benchmark measuring Sana&s image quality compared to leading models – so you are entirely reliant on your own judgment and experiments, including prompt interpretation and coherence. Sana, unlike the best current image AIs, still seems to have the well-known problems with human limbs and the representation of text. You can try Sana (1.6B) for free at generating a 4K image with an Nvidia GTX 3090 takes about 7 seconds here. The source code is available on GitHub: github.com/NVlabs/Sana, and the paper is here: https://arxiv.org/abs/2410.10629. Bild zur Newsmeldung:
 more infos at bei nvlabs.github.iodeutsche Version dieser Seite: Neue quelloffene Nvidia Bild-KI Sana ist 100x schneller als FLUX |
;)
 |
|
|