Must be to a prompt, not an iterative process or a conversation.
Must be consistent, as in, reusing the prompt also works at least 10% (bare minimum) of the time. I will not even consider anything less consistent.
As long as it can generate the English Alphabet keys in the correct order, I will consider it. If the A-Z is correct but the shift key and the caps lock key are swapped, that is counted as well.
For reference, i am refering to this layout:
The prompt must not tell the model the correct layout.
Notably,
neither the poll nor the market predicted the question well, prior to a reliable solution being posted in the comments.
The question has resolved 1 year prior to its close date.
Thanks to all who participated and helped contribute to this question.
@chrisjbillington LOL thank you, I was joking but I'll take it 😂
ARSTDHNEIO for the win! (With due respect for my Colemak-DH using siblings)
@BrunoParga it's a genuinely interesting question! Currently it looks like colmak is much harder for image models.
@chrisjbillington and I would have expected that! It is much less common after all. And since this deals with images: the share of keyboards that look Colemak is smaller than the share that functionally are Colemak. I use this layout and I haven't bothered changing my keys, so if you look at them it's still QWERTY. I suppose those fancy people with fancy mechanical keyboards that have blank keycaps – some of them probably use Colemak as well. So there's less visual training data to begin with.
Try this one on for size @firstuserhere (midjourney, with full credit to @DanaMazalZiv bringing midjourney v6 to our attention for this question)
an image of a minimal computer keyboard on a black background, top-down, full-view, straight-on, simple, accurate, ANSI standard layout, in the style of precisionist, m42 mount, white and beige, duckcore, tumblewave, dutch and flemish, group f/64 --ar 32:15 --v 6.0
Something like six or seven our of twelve with letters correct depending on how much you wanna squint:
@firstuserhere I rate:
1/4
1/4
0/4
0/4
Huh, why are yours so much worse than mine? I posted the first three generations I did, no cherry-picking.
@firstuserhere Ah well, a little more prompt engineering and I'm sure we'll get something that works. I think that's it for me for today though.
@chrisjbillington its still pretty close. I'll try again later in case there's some backend roll out or something.
I don't know anything about midjourney, just signed up for this. Our settings looks the same except I have a lower default model version than you. But --v 6.0
I assume is setting it to use 6.0 on a per-generation basis? I have these options available:
@chrisjbillington these are the models available to me. It'll be very funny if you're using model v5 with a --v6 suffix and it is outperforming my model v6 with a --v6 suffix on this task
@firstuserhere I assume the settings just generate a suffix, and that any arguments you add yourself to the /imagine command take precedent. But i notice --v 6.0 and --v 6 are not the same, it's possible the first one is invalid and I was on 5.2 the whole time? Might check.
@DanaMazalZiv The market criteria say:
As long as it can generate the English Alphabet keys in the correct order, I will consider it.
Post it! That counts!
@DanaMazalZiv Oh wow
About the picture posted:
All the English Alphabet keys are in order (which is what the resolution of the market depends on)
The other punctuation keys are also correct except the keys for "{" and "}"
The top row has all the numbers in the correct order and positions.
The associated symbols are also mostly correct.
Except: "%" and "&"
About the market:
The description states
Must be consistent, as in, reusing the prompt also works at least 10% (bare minimum) of the time
I will test your prompt 10 times and post the results of them here and we can evaluate if we can resolve the market
@chrisjbillington I can reproduce this! The top-left one is correct (w.r.t the the alphabetic keys).
@chrisjbillington it looks like it's about 25%, here's another one I made. I rolled twice, both times I got 1 that got it correctly out of the 4 options
@DanaMazalZiv Nice work, I thought this was very likely to resolve YES, but not so soon. Roughly how consistent is it?
@DanaMazalZiv These are the results I get from your prompt. I've not seen how many of these are correct, yet. The prompt used was:
``` simple vector graphic of a standard QWERTY keyboard --ar 3:2 --v 6.0 ```
@firstuserhere I don't think we have 10% yet, especially since many of these don't show the full keyboard. But with tweaking the prompt to get more full keyboards, we may be able to get it
@firstuserhere try this one: "simple vector graphic of a standard QWERTY keyboard, bird's eye view --ar 3:2 --v 6.0 --style raw"
I ran it 4 times and got 4/16 pictures correct
0/4
0/4
0/4
1/4
0/4
1/4
That's 2/24, not that far off from 10%.
That prompt might already resolve this YES dependent on whether that was an unlucky run