⬇️ Pidgin | ⬇️ ⬇️
English“DemoFusion: Da Breakthrough in AI-Generated Artwork! 🎨”
One big kine problem wit artificial intelligence-made art is da resolution. Even da best kine models, dey stay making just one megapixel (1MP) pictures. Das alright for social media, but if you like print ’em out or make ’em bigger, den you in trouble. Even if you like beef up one real photo wit AI, same problem.
Now, get dis, some sharp minds from da University of Surrey in da UK say dey found one way to make images wit 16 times da resolution of dem uddah big kine players like Midjourney, DALL-E 3, and Stable Diffusion’s SDXL 1.0.
So, wat we get now is dis AI image maker called DemoFusion. Da trick? It keep running da process again and again ’til da quality get bettah. Den, DemoFusion put all da data togeddah. Da best part? You can use ’em for free on one mid-tier gaming computer.
How DemoFusion stack up against oddah AI image makers? DemoFusion, it based on da open source SDXL 1.0 from StabilityAI, one high-performance AI image-making model on top of Stable Diffusion. SDXL 1.0, das max 1024 x 1024 pixels, or 1MP.
Now, Midjourney did some work fo’ make ’em bigger, up to 2048 x 2048 pixels in beta testing. But da base model still da same as DALL-E 3 from OpenAI and SDXL 1.0.
But none of ’em come close to dem smartphone photos. Da latest iPhones and fancy Androids, dey capture images at 48MP and mo’, dat’s at least 8,000 pixels wide, good enough for print. Some Androids, like Samsung Galaxy S23 Ultra and Honor 90, dey go up to 200MP, dat’s ova 14,000 pixels wide. Now, DemoFusion, it get images 16 times dat of SDXL 1.0, dat’s 256MP or 16,000 pixels in each direction.
So, how much DemoFusion go cost? Da folks who make da new model, dey not only make ’em open source but also make sure regular folks can use ’em. Dey share da model and all da good stuff fo’ free so you can download ’em and run ’em on your own computer.
I nevah try DemoFusion on my computer yet, but I wen run one demo version on Replicate, using one Nvidia A100 chip. I wen generate some images up to 13MP. Das plenny good fo’ print ’em out at 300 pixels per inch on one 8 x 12 photo paper, no lose da quality.
Da whole process, from making one image look like Winston Churchill standing on one beach at 13MP, take ’bout 3 minutes. Gonna take bit longah if you run ’em on one gaming rig wit Nvidia GPU or one MacBook wit M1, M2, or M3 chip, but not too much.
So, wat DemoFusion no good at? Da biggest problem, time. Even when you run ’em on da fancy Nvidia chips, gonna take ’bout 10 minutes fo’ make one high-resolution image. Compare dat to seconds fo’ da base SDXL 1.0 or Midjourney.
Gonna be even slower on regular home computer hardware wit gaming-quality Nvidia chips. Professor Yi-Zhe Song, da head of da SketchX AI lab at da University of Surrey, where dey make dis model, he say it wen surprise ’em how good da quality stay, but it take time fo’ make ’em. So, dey thinking ’bout building one new version on top of SDXL Turbo model, dat one start wit 512 x 512 pixel images, half da size of da big kine.
Anotha problem, DemoFusion model, it like fo’ wander and make changes da mo’ times you run ’em. Anyting aftah ’bout nine times da starting resolution, you get big changes from da beginning.
Especially if you use one image fo’ start instead of one text prompt. Dey showed one image of Mr. Bean, and aftah nine run-throughs, he wen look like one whole different person.
DemoFusion, it all ’bout da people. Professor Song say, “Da real goal fo’ us is fo’ get dis in da hands of da people, make artificial intelligence easy fo’ da creative folks.” Next project, dey like give artists da power fo’ fine-tune every part of da image.
Imagine one artist, dey make one rough sketch of like one hand or one bowl of fruit, and den dey use AI fo’ build ’em up piece by piece into one creative artwork. No need just type some words and let da AI do all da work.
I tell you, I find DemoFusion’s quality impressive. I wen see pictures dat look good enough fo’ print, and you can run ’em on your own computer for free. But, gotta admit, if you like da real quick kine, you need one decent gaming GPU, ’cause otherwise, you gotta wait long time fo’ your image.
But DemoFusion, it show us da future. Maybe someday, we go get even bettah image-making in Midjourney and DALL-E.
More from Tom’s Guide: Samsung Galaxy S24 Ultra titanium design just leaked in new real-world images. I just tried Stable Diffusion’s new real-time AI image generator — and it’s scary fast. Outlook for Windows is getting a big upgrade with Copilot AI integration.
Get da latest scoop on da hottest deals plus da daily breaking news, reviews, and helpful tips from da Tom’s Guide crew.
Ryan Morrison, one solid tech journalist, he been doing dis stuff for ova two decades. He rather let his smart articles on artificial intelligence and technology do da talking, not all dis bragging. He da AI Editor for Tom’s Guide, and he use his long time in da industry to explain AI in one way dat make sense, no need be all fancy.
When he not writing his own bio (he no like do ’em, so he let one AI do ’em), Ryan, he into astronomy and physics. He bring dat scientific thinking to his writing. But get dis, he also into storytelling, playing guitar, and even mess around wit indie game making. Dis bio, I made ’em, ChatGPT, ’cause who b.. 🎸💻
NOW IN ENGLISH
High-Res AI Image Generator Beats Stable Diffusion
<p>DemoFusion: The Breakthrough in AI-Generated Artwork! </p><p>One significant issue with artificial intelligence-generated art is the resolution. Even the best models produce just one megapixel (1MP) images. That’s suitable for social media, but if you want to print them or enlarge them, you’ll run into problems. Even if you want to enhance a real photo with AI, you’ll face the same issue.</p><p>Now, here’s the exciting part. Some brilliant minds from the University of Surrey in the UK claim to have found a way to create images with 16 times the resolution of other prominent players like Midjourney, DALL-E 3, and Stable Diffusion’s SDXL 1.0.</p><p>So, what we have now is this AI image maker called DemoFusion. The trick? It keeps running the process repeatedly until the quality improves. Then, DemoFusion compiles all the data. The best part? You can use it for free on a mid-tier gaming computer.</p><p>How does DemoFusion compare to other AI image makers? DemoFusion is based on the open-source SDXL 1.0 from StabilityAI, a high-performance AI image-making model built on top of Stable Diffusion. SDXL 1.0 maxes out at 1024 x 1024 pixels, or 1MP.</p><p>Now, Midjourney has managed to increase the resolution in beta testing, reaching up to 2048 x 2048 pixels. But the base model remains the same as DALL-E 3 from OpenAI and SDXL 1.0.</p><p>However, none of them come close to smartphone photos. The latest iPhones and high-end Androids capture images at 48MP or more, which is at least 8,000 pixels wide, suitable for printing. Some Androids, like the Samsung Galaxy S23 Ultra and Honor 90, even go up to 200MP, which is over 14,000 pixels wide. Now, DemoFusion delivers images 16 times that of SDXL 1.0, which means 256MP or 16,000 pixels in each direction.</p><p>So, how much does DemoFusion cost? The creators of the new model not only made it open source but also ensured that regular folks can use it. They share the model and all the necessary resources for free so that you can download it and run it on your own computer.</p><p>I haven’t tried DemoFusion on my computer yet, but I ran a demo version on Replicate using an Nvidia A100 chip. I generated some images up to 13MP, which is more than sufficient for printing on an 8 x 12 photo paper without losing quality.</p><p>The entire process, from turning an image into something that looks like Winston Churchill standing on a beach at 13MP, takes about 3 minutes. It might take a bit longer if you run it on a gaming rig with an Nvidia GPU or a MacBook with M1, M2, or M3 chips, but not significantly longer.</p><p>So, where does DemoFusion fall short? The biggest drawback is time. Even when running it on high-end Nvidia chips, it takes about 10 minutes to create a high-resolution image. Compare that to seconds for the base SDXL 1.0 or Midjourney.</p><p>It will be even slower on regular home computer hardware with gaming-quality Nvidia chips. Professor Yi-Zhe Song, the head of the SketchX AI lab at the University of Surrey, where this model was created, mentioned that they were surprised by the quality but acknowledged that it takes time to produce. So, they are considering building a new version based on the SDXL Turbo model, starting with 512 x 512 pixel images, half the size of the current version.</p><p>Another issue with the DemoFusion model is that it tends to make significant changes the more times you run it. Beyond about nine times the starting resolution, you’ll notice substantial differences from the original image.</p><p>This is especially noticeable if you use an image as a starting point rather than a text prompt. They showed an image of Mr. Bean, and after nine iterations, he looked like an entirely different person.</p><p>DemoFusion is all about empowering people. Professor Song states, “The real goal for us is to get this into the hands of the people, making artificial intelligence accessible to creative individuals.” Their next project aims to give artists the ability to fine-tune every aspect of an image.</p><p>Imagine an artist creating a rough sketch of a hand or a bowl of fruit and then using AI to build it up piece by piece into a creative artwork. No need to just type some words and let the AI do all the work.</p><p>I must say, I find DemoFusion’s quality impressive. I’ve seen images that are suitable for printing, and you can run it on your own computer for free. However, it’s essential to have a decent gaming GPU if you want quick results; otherwise, you’ll have to wait a while for your high-resolution image.</p><p>But DemoFusion shows us the future. Perhaps someday, we’ll have even better image-making tools in Midjourney and DALL-E.</p><p>More from Tom’s Guide: Samsung Galaxy S24 Ultra titanium design just leaked in new real-world images. I just tried Stable Diffusion’s new real-time AI image generator — and it’s scary fast. Outlook for Windows is getting a big upgrade with Copilot AI integration.</p><p>Get the latest scoop on the hottest deals, plus the daily breaking news, reviews, and helpful tips from the Tom’s Guide team.</p><p>Ryan Morrison, a seasoned tech journalist with over two decades of experience, prefers to let his smart articles on artificial intelligence and technology speak for themselves, without any unnecessary bragging. He serves as the AI Editor for Tom’s Guide, using his extensive industry knowledge to explain AI in a way that makes sense, without the need for excessive jargon.</p><p>When he’s not writing his own bio (he doesn’t like doing it, so he let an AI handle it), Ryan is passionate about astronomy and physics, which he brings into his writing. Surprisingly, he’s also into storytelling, playing the guitar, and even dabbling in indie game development. This bio was created by ChatGPT because who b.. </p>
