Sketch vs. prompt
26 prompts, 26 different caverns. One drawing, the one that was meant.
While building LENS, the team tried to pin down one cavern, Verdigris, in words alone: the tower here, the arches there, the stairs in front. In 13 minutes Mirva made 26 images from prompts of up to 170 words. Each came out a different cavern: the arches, the stairs, the scale and the light shifted from run to run. The final Verdigris was rendered from the artist’s line drawing.
The drawing
The prompts: 26 attempts, 13 minutes
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
“Polished concept art painting, confident economical brushwork, alla prima. An exact composition of a massive underground cavern. Layout: In the bottom left and center foreground, a raised, flat stone walkway with small steps extends into the middle distance… On the left midground, a prominent tall, square ruined stone tower with classical windows stands on a rocky outcrop. On the right foreground and midground, a heavy multi-tiered ruined classical structure with rounded arches…”
One of the prompts, 170 words in full. Real prompts and outputs from the LENS board.
Why words can’t do it
Language is vague about space
“On the left midground” covers a thousand positions. A drawing is one. The model fills the gap with a guess, and the next run guesses again.
Every run starts over
A prompt doesn’t pin anything down, so each image is a new composition. You can’t art-direct a dice roll; you can only reroll it.
The decisions aren’t yours
When the model chooses the layout, the result isn’t your composition. That matters for art direction, and for copyright: a prompt alone generally doesn’t make you the author.
In Mirva, the brush stroke is the prompt
Draw the composition, the silhouette or the pose on a canvas, and direct Mirva from it. Words are still useful for what words are good at: materials, mood, light. The layout comes from your hand, and it stays where you put it.