The Joys and Tribulations of AI-generated Art.

 

AI-art is fun! I have spent more than a year creating AI-images. Despite significant improvements in the algorithms in the past year, the initial images often contain errors. Undesirable features are common and annoying when creating images of people or mythic creatures.

This post illustrates some common problems and describes the fixes for those errors.

Here is an image of a dragon created in NightCafe Studio. Can you see the problems? 

I’ve outlined three errors in the dragon’s anatomy in the next image. 

These types of errors are relatively easy to fix by photo-editing in software such as Photoshop or GIMP (free). The extra leg and pieces of wing can be erased or replaced by background colors. I also brightened the eye. The corrected image is shown below:

This image and other dragons are available as prints or on merchandise like totes, mugs, and T-shirts at https://www.playfulgallery.com/digital-art

And at https://www.redbubble.com/people/auroraspringer/shop

 

You can find more fantastic animals in pictures and poems at my campaign -   https://www.kickstarter.com/projects/auroraspringer/whimsical-creatures-an-art-book-of-fantastic-animals

Follow the campaign to join the fun and help to choose pictures for the Art Book, Whimsical Creatures.  

New Art for Sale - Flowers and a Unicorn

You can find these images and others at my Fine Art Store: https://www.playfulgallery.com/digital-art

And at my RedBubble shop: 

https://www.redbubble.com/people/auroraspringer/shop

Tips for Creating Images from Text with Stable Diffusion and NightCafe

 

The tips in this post are designed to help new users of text-to-images software. I have just over four months of experience in creating images from text. Mostly, I aim to visualize characters and scenes from my Sci-Fi and fantasy stories. I also make pretty pictures for fun.

When I began, I tested different online apps for creating images from text. You can see the results of my initial tests at:

https://auroraspringer.blogspot.com/2022/09/my-tests-of-different-apps-for-creating.html

My tests led me to set up an account with Night Café studio (https://creator.nightcafe.studio/create). I have created more than 1500 images in NightCafe. You can see my published images at https://creator.nightcafe.studio/u/AuroraSpringer

 

Stable Diffusion

To test my text prompts, I often use the free demo version of SD2.1 at https://huggingface.co/spaces/stabilityai/stable-diffusion

(Note, this site is popular and you might need to keep trying several times before you get in the queue.)

Stable diffusion usually requires negative prompts to create people or animals.

Example of the default negative prompt in NightCafe: ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, signature, cut off, draft

(You can edit the prompt as appropriate, e.g., for an image of a bird or a dragon, I might add ‘poorly drawn wings’.)

Advantages of NightCafe (from my perspective)

1. NightCafe provides a variety of methods for creating images.

It offers several different algorithms: Dalle2, Stable Diffusion (I use SD1.5 and 2.1), Coherent (CLIP-guided diffusion), Artistic (VGAN+CLIP; includes animate option to make videos), and Style Transfer.

Each algorithm has advanced options. Dalle has the fewest options, however, Dalle often produces images that more accurately represent what you wanted. 

2. You can evolve from a start image, alter the seed number, and create images in different proportions (square, portrait and landscape) suitable for social media.

3. You can choose from standard modifiers to create different artistic styles.

4. You can upscale and enhance images up to 8000x8000 for Print on Demand applications. (Important since many sites do not include this option.)

5. You choose whether to make an image public or keep it private.

6. You can use the images for commercial purposes even on the free plan.

7. The user interface is straightforward, and the user community is friendly. Members of the FaceBook group NightCafe AI Art offer advice on stable diffusion, tutorials and example prompts.  

The administrators are willing to adopt suggestions for improvements.

8. NightCafe has a flexible pricing model (Important if you wish to avoid a monthly fee.)

Each image costs 0.5 or more credits. Most methods are 1 credit to create a panel of 4 images. Dalle2 is 3 credits for a panel of 4 images, and SD2.1 is 4 credits per 4 images.

You can subscribe with a monthly fee, purchase credits, or earn credits.

You get 5 credits every day, and can add 2 more by voting in the daily challenges. You can earn credits by sharing an image and acknowledging @NightCafeStudio on Twitter, Instagram, and Tiktok. This works well for me since I routinely post images related to my books on social media.

You can check the best of my images at my site for WhimsicalArt. These images are available as prints, magnets, T-shirts and other merchandise.  

(Most of these images are unaltered, although some required polishing with GIMP or Photoshop.)

I write science fiction and fantasy adventures entwined with romance. You can find a list of my published stories HERE.

Explore my Blog for background and excerpts.

I post regularly on Facebook,  Twitter,  YouTube,  Instagram  and TikTok