Accomplice June 2022 updates: DALL·E Mini, 11 other new models, and you can now use a prompt as a starter image

Over the last 48 hours the amazing versatility of DALL·E Mini has gone viral and I'm excited to announce that it is now a new default Accomplice model!


Over the last 48 hours the amazing versatility of DALL·E Mini has gone viral and I'm excited to announce that it is now a new default Accomplice model!

DALL·E Mini is the largest open-source AI image model yet – created by Boris Dayma and based off of OpenAI’s DALL·E and DALL·E 2 – and just like all the other text-to-image models Accomplice supports, it generate images from any text prompt you give it.

But, what makes DALL·E Mini really special is – because of how big it is – it knows an astounding amount of stuff about the world, and more often than not does a suprisingly good job of combining animals, places, pop culture references, and artists' styles all into one coherent image.

It's a lot of fun, extremely useful, and on top of all that – it's still a work in progress! DALL·E Mini is still being actively trained and will continue to improve and only get better and better.

More new default models

New default models Background Remover and Super Resolution

For several years two of my favorite AI-powered image tools have been and The other day – while using both of them for the 100th time – it finally dawned on me that Accomplice already had all the pieces necessary to create models that would replicate them, and that they – more importantly – would be a useful addition to the Accomplice toolbelt.

So, now, the new “utility model” Background Remover will remove the background from any image that you've either created or uploaded.

And Super Resolution will increase the size of any image you've created or uploaded – as long as its under 5000px – by at least 2x. So next time you wish an image you created with Accomplice could get just a little bit bigger? Run it through Super Resolution.

New style transfer model

Style Transfer is one of Accomplice’s most popular models and the new Anime Illustrator model is a great new addition to the style transfer model family.

It knows human faces and does a great job accentuating features. But it can also do places, animals, and things and works great as just a generic and versatile illustrated style transfer.

8 new premium models

Last week I went a little crazy and trained 6 new premium models – using Accomplice’s super easy drag-and-drop no-code model builder – on the works of my favorite abstract artists: Pollock, Rothko, Kandinsky, Mondrian, Klimt, and Miró.

While Accomplice has always been good at reproducing abstract artists’ styles, these specifically trained models are a bit higher quality and more accurate than what you would get from ImageNet or Disco Diffusion because they have been fine-tuned on the specific artist's portfolio.

Plus, since most of these models can use the new, larger aspect ratios that custom trained models now support, you can create landscape, profile, panoramic and skyscraper sized images!

In addition, after watching Tim Burton's Big Eyes, I got inspired to train a model – Big Eyes AI – based on Margaret Keane’s iconic big eyed children style.

And, last but not least, I created a model I'd been wanting for awhile now – Woodgrain Background AI. Perfect for the next time you need a nice, flat, knotty image of pine, cedar, oak, or the like for the background of a sales deck or webpage hero.

New: Use a prompt as a starter image

It's now easier than ever to chain models together by taking something you created using one model and using it as either:

  1. the starter image prompt for ImageNet, WikiArt, or Disco Diffusion, or
  2. the content image for Style Transfer or Anime Illustrator, or
  3. the image input for the new utility models Background Remover and Super Resolution

Now, when you go to add a starter image prompt, you'll see a new "Use a prompt" button below the "Upload an image" button. Selecting that button will take you to a screen full of all your most recent created images. Choose one of those prompts and it will use the main thumbnail image from that prompt as the starting image for whatever model you want to create with.