It all started with DALL•E

 kw: photo essays, art generation, generated art, history

Four years ago I first got access to DALL•E, version 2. Version 1 was almost two years old already when Version 2 was released by OpenAI in September 2022 and I began to use it in November. Based on only the briefest of introductions, I created a prompt that I thought would be a bit of a challenge:

Mountainous Landscape in the style of the Hudson School

Working only from memory, I left out the word "Valley" for the Hudson Valley School of landscape painting. Nonetheless, the art engine produced four images with the appropriate "look" and presented me with these choices:

These images were 1,024 x 1,024 pixels, and I learned that prior art generators all topped out at 512 x 512 or smaller. The set of colored blocks in each image is the logo of DALL-E 2 (I'll use a hyphen in the name hereafter rather than the bullet). The image I like best, that looked the most like a HVS painting, is the second one:

OpenAI was touting Outpainting as a new thing in November 2022, so I tried it on this image. Here is the result, which used up most of my trial credits:

This image size is 1,920 x 1,344. Once I began post-processing generated images for use as wallpaper, I cropped this one to 1,920 x 1,080. That discarded the bottom 20% of the image, but it had the benefit of removing the logo also. From this time forward, when I've made a wallpaper file I have "signed" every image with my signature and the name of the art generator. IrfanView has a useful Paint function for this.

I soon began buying credits to use DALL-E 2. I also heard about MidJourney from my brother, and Stable Diffusion in a blog post by a friend. I tried each of these. At the time MidJourney could only be used through Discord; somehow I could never get used to using Discord and I haven't tried since early 2023. The Stable Diffusion interface was also harder to use than DALL-E 2, so I tried a couple of prompts and didn't try again. DE2, as I began to call it, kept me contentedly busy for about a year. Nearly eighty of the images I produced with DE2 pleased me enough that I produced wallpaper versions, almost all with an exact 16:9 aspect ratio, 1,920 x 1,080 or larger.

At the end of 2022 I learned that DE2 was being phased out, and that its replacement would be DE3, available through the Bing Image Generator in the Edge browser. However, DE3 didn't have Outpainting, but until DE2 became unavailable I would use DE3 to produce a higher-quality beginning image, and outpaint it in DE2.

Over time, as I learned of new art generators, I would try them. Some came, lasted a while, and were discontinued. One I miss is Playground, because it was the first engine aggregator I came across. I could use several versions of Stable Diffusion and a dozen or more others (most of which I've forgotten), including Playground's own engines. There were dozens of "filters" for different kinds of images, such as drawings or acrylic paintings or cinematic views. But by the end of 2024 the company changed strategy and focused on graphic design, dropping art generation.

Skipping boring detail, we can fast-forward to the present. I subscribe to two aggregators, OpenArt AI and Leonardo AI, which each seem to provide access to about 100 art engines, as I call them. New ones are added from time to time. I also use the free versions of Gemini, Google Flow, two versions of Microsoft AI Image in Bing Image Creator, and on occasion whichever GPT engine is available at the moment in DuckDuckGo.

During the past week I noticed that OpenArt had released four new engines, each called a "flagship". I decided to try them with the short prompt I used four years ago, with the addition of the word Valley in the prompt. All but a very few of the OpenArt engines can create images with a 16:9 aspect ratio (approximately, in some cases) and some can make them 1k or 2k and sometimes 4k. Depending on the cost of images, I might create all 2k images, or 1k for later upscaling with Upscayl, which I obtained in mid-2024. There follow the four engines' productions, the 4-image preview montage and then the image I like best from each montage.

First, GPT Image 2.6 Flare (faster):

When I can I avoid having the sun visible in the image, my primary reason for choosing this image of the four.

Second, GPT Image 2.6 Sunburst (more detailed):

Here, all four had the sun in the frame. This has the most pleasing arrangement, in my view. This image and the prior one are 1,792 x 1,008, or 112 x 63 times 16 x 16. I've noticed that all art generators work on 16 x 16 units. To render these images for wallpaper I will upscale them 2x to 3,584 x 2,016. No further cropping is needed because this is an exact 16:9 aspect ratio. If I later want to use them with a 4k screen, I'll re-upscale the originals 3x, to 5,376 x 3,024.

Third, Seedream 5.5 (from ByteDance):

Image size is 1,360 x 768 (85 x 48 times 16 x 16). For wallpaper I'll upscale to 2,720 x 1,536 and crop to 2,720 x 1,530.

Finally, Nano Banana 2.1 (Google):

If you look carefully you'll see that all four previews have at least one person in them. Nano Banana really "likes" having people in its images. Image size is 1,376 x 768 (86 x 48 times 16 x 16). Upscaling 2x produces 2,752 x 1,536. The way I like to process images this size is by rescaling in IrfanView to 2,741 x 1,530, then cropping to 2,720 x 1,530.

When I want a 16:9 wallpaper image, I aim for a multiple of 32 x 18.

All these images are more detailed than the one DALL-E2 produced. I have a certain affection for that DE2 image, though, because it is my first foray into the world of art generation. All my life I've loved art, but I cannot draw well. For a long time I tried and tried. Now it is like having a "stable" of artists I can hire economically, and some work for free. I get the ideas, they do the painting/drawing. Each image I produce is a "negotiation" with one of these artists.

In four years I have amassed more than 2,000 wallpaper files in three dozen categories of subject. From time to time I hope to present selected images from various categories…as time and interest dictate.