Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[dragdeler]

The flame is a camelid from the andes. It has been domesticated since possibly hundred of thousands of years and is commonly agreed to be the spark that lit civilatory progress. Don't look into it's eyes or it will stare right into your soul.

Spoiler (click to show/hide)

10 out of 10

[voliol]

Behold! A man!

Spoiler (click to show/hide)

It works strangely well on "dog", "mushroom", "lamp" and "bird" though. Iirc those earliest dreaming AIs were only trained on e.g. those objects, creating those freaky landscapes made out of warped dogs and birds.

[Naturegirl1999]

Maybe the trainers heard of Diogenes

[Starver]

Just come across this.

Won't bother with the pictures (it appears you can probably conjure something up yourself, anyway...) just go with the input and my description of the output:

"Black and white house" -> some type of grouse? A bird, mostly black, but blue throat highlight and red crest.

"Jigsaw puzzle" -> Strange strange image... charitably with 'jigsawish cut patterns' echoed in the texture.

"Freight train" -> Looks like a collage of things that don't look at all like trains, but composed surprisingly like a side-on shot of the link between a couple of carriages/trucks on an ebankment track.

"Blue string soup" -> A composite picture of several measuring jugs?

"'Allo 'Allo" -> a distorted picture of some perched eagle, or similar.  It's head much 'overfeathered' and hardly any beak or obvious eyes visible.

"Cripple Mister Onion" -> Some blue (colour-shifted?) heron-like wading bird, with one of its legs overly distorted (or not fully imported onto the background.

So I see what some of these might be doing. What's missing is the feedback. Obviously it relies upon a separate (internal/otherwise privided) deep-learning to even get the results it does.

But I'd appreciate some form of feedback. Even if it gets largely ignored, only lumped into general algorithmic training if one agrees with a sufficient number of peers (less any sufficient number of counter-opinions). As it is, I get no feeling of accomplishment. Though I'm just a few steps from Keyboard Mashing total nonsense (as opposed to what I actually tried) to see what it gives with "fhfvhnnhfvdsn fdsnfcoihs" or somesuch.

...Ok, so I did. That turns out to be a small perching bird on a twig - perhaps a finch (or mostly a combination of finches) with a rose 'bib' and scarlet 'cap' (either HSV-shifted or imprted from another species-picture) but other plumage being cream/dark-brown.  Repeating it (the first time I've done that with a submission) gives a different view of very much the same bird but with slightly blue cheeks and blue-green tail-flashing. Too close to be sexual dimorphic, too different to make me think it's just a gender/age difference. (If it were a real bird.)  So uncanny valley.

And a lot of the 'training'/source pictures were from ornithology resources.  Far more than the landscape/architecture/machinery ones that might have seeded the ones that were not birds, in my very brief experiment.

Fun for a few minutes, definitely. I've now got an inkling of what sort of mashup is going on at the backend. Not addictive enough to keep playing with, though.

[Naturegirl1999]

Spongebob Squarepants gave me this

[Starver]

Ok, so I played more, to see if I could try to play towards its bird biases.

"Sea Eagle"... nearly.  "Golden Eagle"... well, a bird, but not really the right one.  "Osprey" seems to be some sort of fluke worm ('prey'?).  "Condor"... Obviously trying to show me a conch shell...

It's good with "Salamander" and "Newt" and other amphibian names.

"Factory" gives me something obviously based upon a tractor ("...actor..."). "Tractor" gives tractor(s) in interesting takes.

"Bloodhound"/"Dalmation"/"Pug" work well, so it knows dog breeds too (though I never obviously got dogs for anything else). "Rover" gives a dog (Boxerish, more so than "Boxer", the one time I tried that). But "Fido" defaults to furniture/room-fittings.  (Leading me to then try common furniture names "Table", "Chair", that all seem to go well in that direction, though the chair seemed to be part-morphed with a person, head-as-headrest, at least one arm as an armrest... maybe that person was sitting on one of the source-image chairs?)

...I can also get interesting landscapes if I clearly suggest something like "Cliff".



(But move away from the obvious, i.e. into a whole in its knowledge, and it gets quite entertaining.  As "Starver", I am a beaver.  Confirmed with a few submissions, though the way the first one was cooked up actually looked more like a Duck-Billed Platypus.)

[voliol]

”Spongebob squarepants” gives the same kind of results as ”man”. For ”man” perhaps because of sala”man”der?

Trying ”voliol” I get strange-looking lions, with squished faces. Does the face-squishing come from some other word it’s recognizing in ”voliol”? ”Vole” maybe? Nope, a vole is a pole.

I get the most cursed images with ”a cat and a dog”, composed of human skin/fingers, some red and white patterns and sometimes nails, in unexplicable shapes. ”A dog and a cat” returns lemurs. ”Cat” is some kind of (well rendered!) terrier, ”a cat and a cat” feline amalgamations.

[Kagus]

In an earlier version of Dwarf Fortress, players were mesmerized and confused when observing that their dwarves would regularly excavate and haul away a "Large ", with no idea of what such a thing might be.

At last, we have the answer.

Spoiler: Large (click to show/hide)

EDIT: I also find it highly amusing that "Penis" primarily returns images of guys holding up fish. Dating website profiles the world over have finally been explained.

[Bumber]

Spoiler: Large (click to show/hide)

Image doesn't work. 404 error.

[Naturegirl1999]

I remember what it was, sort of, either a kangaroo or a very large rabbit, it was facing the screen

[Kagus]

Just a rabbit with an adorably/horribly messed-up face. I've seen a couple pop up with bird feet, which was a curious sight...

[Starver]

I noted this when I spotted the already old thread. The screenshots/save+host would be more permanent records, unless you don't mind just regenerating just roughly the same thing anew (so long as they don't update the compilation data beyond a certain point).

I suspect that it caches (and expires out after some period/interim new cache volume) rather than regenerates from the same hash each time[1]). Though I suppose you could submit the URI as a forum avatar to see if regular revisitting by random passers-by in that community sustains it. (Bad netiquette, I'd say, in several ways, but could help establish the exact retention mechanism if it bothered anyone enough to want to find out.)


I think its main issue with heads/faces (of various kinds) is that its 'learnt' composition rules clash with (or totally ignore) our own inbuilt pareidolia, thus creating an "uncanny gulch" where other bits of the composite body/scene might be largely 'passable'. That said, I recall that Large having rather oddly proportioned front legs, too. Or maybe I just don't know the right kinds of lapinate breeds to recognise an accurate example.


[1] Probably originally sequentially-/time-salted, given you can get different ones generated for successive inputs. Didn't look into the API options, just the main submission page clicky thing. There might be a couple of interesting ways to probe this. If one were bothered enough.

[Bralbaard]

Filling in "Urist" consistently generates engraved slabs with unintelligible arcane writing.



I'm fairly certain some of these contain the secrets of life and death.

[Loud Whispers]

Behold! A man!

Spoiler (click to show/hide)

It works strangely well on "dog", "mushroom", "lamp" and "bird" though. Iirc those earliest dreaming AIs were only trained on e.g. those objects, creating those freaky landscapes made out of warped dogs and birds.

Lmao Diogenes would love this