Re-visiting the Jabberwock

 kw: ai experiments, artificial intelligence, generated art, poems, illustrations, photo essays

A little more than two years ago I used Dall-E2 to illustrate the poem "Jabberwocky" by Lewis Carroll. Now that there are four high quality generative art sites that I use, I experimented with bringing just the Jabberwock itself up to date. In nearly all cases the programs relied on dragon imagery.

This is a wide-view redraw by Dall-E3 of one result of my second prompt. It's a little less sinister than the results of the first prompt. Here is how it went:

Dall-E3: "The Jabberwock"

All are fierce dragons in near-silhouette. All but one are shown against the Moon.

Dall-E3: "The Jabberwock as a pastel painting"

The word "pastel" has evoked a less sinister atmosphere. We find flowers and butterflies and in one case the dragon is actually smiling. Now that Dall-E3 can recast a square image into one with a wider aspect ratio I had it re-do the second image, on upper right. Note that the wide image at the top of the post has elements from all four images. By experimentation I found that, while re-formatting each image produces a different result, repeating the re-format step on a particular square image yields exactly the same wider image as the prior re-formatting on that image. The re-formatting offerings are square and 4:3, but the resulting wide format images are actually 1792x1024, with a ratio of 1.75:1 or 7:4. That's close enough to 16:9 (HD) that it takes just a little cropping to produce a 16:9 image (1792x1008), which can be put through Upscayl to make it big enough to use for an HD wallpaper.

I went on to try Gemini (formerly Bard), which now uses Imagen 3 to make images, and only one at a time. The 3-panel pasteup here is from three prompts:

• An image of the Jabberwock from the poem Jabberwocky
• An image of the Jabberwock from the poem Jabberwocky as a pastel painting
• A full-body image of the Jabberwock from the poem Jabberwocky as a pastel painting

Adding "pastel" definitely produces a less fierce result. The third item is the least dragonlike. Unfortunately for much of what I do, one cannot outpaint with Gemini, and asking for "wide format" or "HD ratio" gets ignored.. Maybe such options are available with Gemini Plus, but I haven't yet signed up for that.

I looked next to Leonardo AI, which is the most recent tool I've used. It has many options and variations. Here I'll present four sets of results. Leo produces images in a horizontal string. I rearranged these as 2x2 rectangular arrays. The original images are 1368x768 pixels, not quite a 16:9 ratio (though it's labeled as such in the menu), but 171:96.

Leonardo: "The Jabberwock from the poem Jabberwocky as a pastel painting", Portrait, Cinematic

The "eye" in each image is a flag that indicates it is public. One needs a paid subscription to make private images. Though these are all dragonoid, some of them have no wings. This prompt without "pastel painting" produced much darker and fiercer results.

Leonardo: "The Jabberwock from the poem Jabberwocky", Concept Art, Stylistic Illustration

These were the most colorful results of all my experiments with Leo. The "Concept Art" setting pushes such limits. Though these have fierce expressions, they are rather cute.

Leonardo: "The Jabberwock from the poem Jabberwocky", Graphic Design, 3D

These are quite dark and sinister, but not as dark as some results.

Leonardo: "The Jabberwock from the poem Jabberwocky as a pastel painting", Graphic Design, 3D

The addition of "pastel" has lightened things up a lot, both in visual and emotional tone. It's interesting to note that these critters all have multiple tails. I count lavender as the least threatening hue.

Finally, I turned to Playground, which was once the most flexible of the tools. Since they shut down Canvas mode in mid-September, it has been harder to use all the filters and other options. The free version is slow, and sometimes times out. Playground also presents a horizontal string of four images, which I rearranged as a 2x2 matrix. I set aspect ratio to 4:3 mode, which one would expect to yield 1.333:1 images; the actual images are 1216x832, about 1.46:1 or 19:13. For the two Playground mode where this was possible I used the Watercolor filter, considering it similar to saying "pastel painting". Playground has three image engines available.

Playground: "The Jabberwock from the poem Jabberwocky", Stable Diffusion XL, Watercolor filter

The background architecture in items 1 and 4 adds interest. SDXL is now considered a "traditional" or even "retro" engine. I also had the option to set a "faithfulness to prompt" setting, and used 4 ("more free").

Playground: "The Jabberwock from the poem Jabberwocky", Playground v2.5, Watercolor filter

As I've frequently observed, PG25 is edgier and typically a bit darker.

Playground: "The Jabberwock from the poem Jabberwocky", Playground v3.0, no filter

Wow! What a difference from all the others. PG30 is apparently more attuned to "poem". It even tried to write one (second image), though the character strings are illegible. The third image looks like it could be a bookplate. PG30 is now the most colorful and creative engine Playground has to offer; it was introduced earlier this year.

I must say a word about limits and conditions in the free versions I use.

• Dall-E3 lets you run 15 prompts daily, and each prompt yields four images. There appears to be no charge for re-formatting an image.
• Gemini has no explicit limit, but when I asked, it told me that sending many image requests in rapid sequence could slow things down.
• Leonardo AI gives you 150 points to use in a day. However, that doesn't mean 150 images. There are about ten (the number occasionally varies) Presets. Two of them "charge" 10 points per prompt, and in the free version, you always get four images per prompt. One of them is 24 points (on a few occasions I saw it was 104 points). The rest are 14 points. So it is possible to generate fifteen 10-pointers, or ten 14-pointers plus one 10-pointer, or six 24-pointers (with 6 unusable points) per day. A "day" resets to the time you created your account, which is 8PM for me. You can outpaint and inpaint with Leonardo, for variable numbers of points per action.
• Playground originally let me generate 150 images daily, up to four at a time. More recently the limit has been 50, and at present I don't see an indication of how many images I have left in a session.

Paying for a subscription to Playground or Leonardo AI opens up greatly expanded limits, and added functionality. I haven't tried all the combinations of either of these programs, but one day I may pick one of them to subscribe to. I have lots else going on in my life, so I don't see much need to do so at the moment.