AI Interpretations of Jane Eyre’s Paintings

In chapter XIII of Jane Eyre, Mr. Rochester asks to see Jane’s painting portfolio. In an act of 19th century negging, Mr. Rochester identifies three of her paintings as most worthy:

…you have secured the shadow of your thought; but no more, probably. You had not enough of the artist’s skill and science to give it full being: yet the drawings are, for a school girl, peculiar.

The three paintings are described quite explicitly, so I gave their descriptions to Midjourney and Dall-E 2 to see what would happen. I originally envisioned this post being a kind of contest between the two AI art generators, but Midjourney would have won by such a large margin that it would be pointless.

The three descriptions are as follows:

The first represented clouds low and livid, rolling over a swollen sea: all the distance was in eclipse; so, too, was the foreground; or rather, the nearest billows, for there was no land. One gleam of light lifted into relief a half-submerged mast, on which sat a cormorant, dark and large, with wings flecked with foam; its beak held a gold bracelet, set with gems, that I had touched with as brilliant tints as my palette could yield, and as glittering distinctness as my pencil could impart. Sinking below the bird and mast, a drowned corse glanced through the green water; a fair arm was the only limb clearly visible, whence the bracelet had been washed or torn.

The second picture contained for foreground only the dim peak of a hill, with grass and some leaves slanted as if by a breeze. Beyond and above spread an expanse of sky, dark-blue as at twilight: rising into the sky, was a woman’s shape to the bust, pourtrayed in tints as dusk and soft as I could combine. The dim forehead was crowned with a star; the lineaments below were seen as through the suffusion of vapour; the eyes shone dark and wild; the hair streamed shadowy, like a beamless cloud torn by storm or by electric travail. On the neck lay a pale reflection like moonlight; the same faint lustre touched the train of thin clouds from which rose and bowed this vision of the Evening Star.

The third showed the pinnacle of an iceberg piercing a polar winter sky: a muster of northern lights reared their dim lances, close serried, along the horizon. Throwing these into distance, rose, in the foreground, a head,—a colossal head, inclined towards the iceberg, and resting against it. Two thin hands, joined under the forehead, and supporting it, drew up before the lower features a sable veil; a brow quite bloodless, white as bone, and. an eye hollow and fixed, blank of meaning but for the glassiness of despair, alone were visible. Above the temples, amidst wreathed turban folds of black drapery, vague in its character and consistency as cloud, gleamed a ring of white flame, gemmed with sparkles of a more lurid tinge. This pale crescent was “The likeness of a Kingly Crown;” what it diademed was “the shape which shape had none.”

I’ll note that prior to their descriptions, the paintings are noted to be watercolors.

Midjourney: Artistic, but inexact

For a first pass, I tried the prompts as-is, with the word “watercolor” at the very start.

While these all get the general idea, they miss a lot of the details. For example, the first completely misses the drowned corpse and the fact that the ship has sunk; the second misses the star crown; the third misses the fact that the iceberg and the head are not the same thing. And lots more, too.

For my second attempt, I rewrote the prompts to try to encourage the AI to hew as close to the description as possible. This didn’t really work, but it produced some interesting art:

water color. the mast of a sunken ship, cormorant holding a bracelet, drowned body in the water, roiling clouds for the first painting generated

water color. girl on a hill with a star on her forehead, dark colors, leaves, her hair is blown by the wind, so are the clouds for the second generated

water color. a giant head rests against an iceberg. northern lights. the head wears a crown of white fire for the third generated

Midjourney is great at producing images that look like art, but it’s not good at parsing the details of the prompt. It seems to care most about words near the start of the prompt. For example, if I moved “watercolor” to the end of the prompt, it ignored it entirely.

Dall-E 2: Inexact and uggo

Unfortunately, the full prompts are too long for Dall-E, so I just used the same edited versions I’d made for the second attempt.

Note that Dall-E, like Midjourney, gives four different outputs. Sadly, you can’t save them in one convenient file, so I just picked the nicest looking ones. I don’t think any of the Dall-E outputs compete with their Midjourney counterparts.