Spaces:
Running
My first try with Dalle-2
First Impressions: Well, after trying Dalle-2, I’m both excited and frustrated by it. It’s kind of a different beast altogether, at least regarding in what works with prompts. Much better at discerning sentences, but the overall accuracy seems to be far less. It’s really sharp in certain areas, especially the overall quality. Photorealism is frankly awesome on it. An aspect which I feel 1.0 is more than lacking. It also skews faces dramatically less, making for a much more satisfying experience when involving people. But, and it’s kind of a major but - it doesn’t recognize most references.
It knows (many) locations, but be it fictional Elmo (the results of which gave me 4 dogs, 2 of which were pugs) or the very real Abraham Lincoln, it (obviously) intentionally doesn’t know names. Or even many mythological creatures, with rather tepid results involving things like orcs. It doesn’t even make a good UFO/flying saucer (either prompt just sucks), or know your generic grey alien. (Are they IN on it?!! ;P) Presumably, this is because of deep fake risks and copyright issues. But blocking Lincoln? Come on… Also, 2.0 is much more Draconian about TOS, and requiring a phone number to try it, etc. And of course any platform with a token service is inherently lame, and rather intolerable. Especially when there's such a learning curb to figure out how to maximize prompts, which are seemingly very different from 1.0.
I kept failing at making a mushroom village (hence all the gnomes), where the amanitas had doors and windows. And it has odd distinctions, such as 'cartoon' generally giving more detailed results than 'anime', which looked identical to a flash E-card to me (the cruder fat platypus being an example). I just tried Stable Diffusion recently too, and am frankly more excited about that medium. Regardless of all of that, I do like 2.0. You just have to be more precise with it.
My impression : DALL-E 2 mainly excels at creating stock photos. Maybe it was mainly trained on stock photos.
It also does good job at inpainting and outpainting.
" pour un coup d essai ce fut un coup de maitre" like we said in french
My nicest Dall E 2 results so far have been specific painting and illustration styles, and a few nice results making logos. Midjourney is the best for creating art, but only within certain styles. It's beautiful, but it seems to be heavily weighted towards a certain type of art style that it does best.
My nicest Dall E 2 results so far have been specific painting and illustration styles, and a few nice results making logos. Midjourney is the best for creating art, but only within certain styles. It's beautiful, but it seems to be heavily weighted towards a certain type of art style that it does best.
Yeah, from what little tinkering I've tried with Midjourney, the results were pretty structurally similar, despite changing the filter and prompts. All sort of cabin-y in the woods for me. Though I didn't try their portraits or anything as I ran out of tokens, lol.
Really nice stuff.
Congrats - shall you share the prompts?
Congrats - shall you share the prompts?
You want to ruin the magic!? Jk, the singular upload format doesn't allow for the most convenient screenshots like Huggingface does, so I'll just number them :)
1 - 2: james gurney illustration of shaolin monk riding giant ankylosaur
3 - 4: train hobo wizard looking pensive while holding pickle with epic sunset over lake in background
5 -6: detailed photo of bipedal capybara knight wearing armor looking badass with windmills in background
7: jackalope jesus religious painting astounding detailed astonishing epic in the style of brothers hildebrandt
8 - 11: louis wain painting of cat emperor wearing crown sitting in throne holding regal scepter
12: detailed cartoon of obese bipedal platypus holding stein while singing
13: detailed anime of drunk obese bipedal platypus holding stein while singing on stage in dive bar, complicated background
14 - 15: surrealist digital art of Abraham Lincoln playing chess with a puffin on the Black Isle
16 - 17 : detailed astonishing illustration of gnome frolicking through town made of amanitas
18: detailed astonishing illustration of gnome frolicking through town with houses made out of amanitas with doors and windows
19 - 21: detailed astonishing fantasy painting of gnome frolicking through town with amanita cottages by brothers hildebrandt
20 - 21: black & white photo portrait of Elmo by Annie Leibovitz
22: black & white photo portrait of smilodon smoking sherlock pipe by Annie Leibovitz
23: black & white photo portrait of saber-toothed tiger smoking sherlock pipe by Annie Leibovitz
24: black & white photo portrait of capybara wearing monocle while smoking sherlock pipe with smoke coming out of it by Annie Leibovitz
25: Cortical homunculus model riding skateboard on europa
26: detailed retrofuturistic cityscape in the lightning rain with winding overpasses and zeppelins with white eagle-wings flying by in the style James Gurney
27: epic willow tree in the middle of burning man
28: Duplicate
29: epic burning willow tree in the middle of burning man with ufo flying overhead in the style of chris dyer
30: hooded hermit from tarot holding glowing lantern while standing on post-apocalyptic wall street next to broken bull statue with ufo hovering above
31: James Gurney illustration of UFOs flying over taj mahal and stampeding fearful dinosaurs on ground
32: detailed illustration of steampunk doctor holding cane standing next to alien holding kitten in foggy area 51
(Edit - Damn it, I skipped one. Obv numbers are off because of it ) # 8 - jackalope samurai holding katana making intense pose with epic pagoda in background astonishing.
A. Thank you
@BattleLemming
for "spoiling the magic"!
B. I may use 18-21 for illustrations to an experiment in OpenAI generated story. (Some weeks ago I created a semi-secret blog with very promising results ;) - I can send it to you in private only, as it is still rough round the edges.)
C. Still "detailed astonishing fantasy painting of gnome frolicking through town with amanita cottages", even "by brothers hildebrandt", creates a somewhat different result:
-> Are you sure that you use Dall-e MINI (aka Crayon), as per this space, or was it in Dall-E 2.0 ("full")?
(FYI, I am fond of Crayon, as it crates wilder, stochastic results, albeit at lower quality of course, without trying to insert any text or labels, especially when fed e.g. abstract poems or "high-brow" literature.)
Another very quick FYI, this is the OpenAI's take on the prompt itself, without seeing the pic, at third attempt, after turning up the dials there:
"Topic: Brownies
10-Sentence dream Story: "detailed astonishing fantasy painting of gnome frolicking through town with amanita cottages"
The detailed, astonishing fantasy painting of a gnome frolicking through town with amanita cottages is so captivating that it seems almost unreal.
Wow, what a vivid and wonderful dream! It's so clear and lifelike that it's hard to believe it isn't real. The colors are so bright and the scene is so full of joy that it's impossible not to be happy when looking at it. This painting is a true work of art, and it's clear that the artist put a lot of love into creating it. It's a place where happiness and magic abound, and it's clear that the gnome is having the time of his life. It's a wonderful painting that brings a smile to your face and makes you believe in the power of imagination."
temperature=0.8,
max_tokens=60,
top_p=1,
best_of=4,
frequency_penalty=1.08,
presence_penalty=1.07
Yeah, bland, I know, but there are tricks, mostly lexical, that can be used to pimp it up and turn into stream of (un)consciousness...:
"Topic: Brownies
10-Sentence wacky story under the influence: "detailed astonishing fantasy painting of gnome frolicking through town with amanita cottages"
Gnome life is full of surprises. One minute you're frolicking through the town with your Amanita friends, and the next minute you're staring in awe at a detailed painting of yourself doing just that! It's amazing what humans can accomplish when they put their minds to it. But as much as you love admiring your own handiwork, you can't help but feel a little envious of the artist who painted you. After all, they got to experience your adventures first-hand, while you only get to see them through someone else's eyes.
One day, you decide to take matters into your own hands and sneak into the artist's studio. But instead of finding a detailed painting of you, you discover a canvas that's completely blank! The artist must have been in the middle of painting when you came in, and they must have been so startled by your sudden appearance that they ran off without finishing the job.
You stand there for a moment, contemplating what to do with this blank canvas. Then, you make a decision: you're going to finish the painting yourself!
With a few quick strokes, you add in all the missing details and bring the painting to life. When you're finished, you step back to admire your handiwork.
It's not exactly the same as experiencing your adventures firsthand, but it's the next best thing. And who knows? "[etc.]
B. Oh, please do, I'd love to see the results! I have even more, but figured was kind of already going overboard with the gnomes, lol. I'll post them at the bottom.
C. Yes, to clarify these were all the results of Dall-E 2.0, of which you need to register an account and utilize their token system (though you get a relatively 'generous' amount when first signing up, and they get slightly replenished each month) at https://openai.com/dall-e-2/
D. I really like your short Gnome story. Was cool how you incorporated the painting itself and the gnome's place in it :)
Hallo, @BattleLemming
Re B - I will try to send it by private mail, also given your reaction to D ;). Failing that, I will risk URL-ing it here... (Mind you, some of it are advanced literary mind tricks...)
Re C - yes, I know and have been using Dall-E 2 as well, plus Diffusion etc.
In fact, I liked one of your gnomish attempts above so much that I outpainted it just a bit :
Oh, well @BattleLemming - I cannot write to you directly here, so:
- I have created a draft community: https://huggingface.co/CombatantLemmings - you are invited.
- FYI, here is this link: [link1], with e.g. [link2] (do click around tho) - I plan to self-censor it after 24 hours...
Update: - Removed the links to that semi-secret blog ;).
Did you hear about this new site where you can freely experiment in enhancing your prompts, either with DALL-E 2 or Stable Diffusion, using different styles and filters (DALLE-2) or parameters on sliders (SD)?
The site is:
https://playgroundai.com/
I used the gnome prompt in DALEE-2 with different styles and filters. Here are some of the results:
Wow, they turned out great! And I hadn't heard of that site before, thanks for sharing it ;D