Adventures in Dall-E

Best Selling RPGs - Available Now @ DriveThruRPG.com

robertsconley

Legendary Pubber
Joined
May 3, 2018
Messages
5,585
Reaction score
14,468
Dall-E is an AI server that can generate images from English language descriptions. It can also generate variations of existing images.

1661990624906.png


1661990650909.png

a tall muscular orc warrior wearing chainmail wielding a halberd as black and white line art

1661990677699.png

a medieval wooden Russian fort within a evergreen forest as black and white line art

1661990723787.png

a medieval Russian village within a evergreen forest as black and white line art

1661990773502.png
 
I had a go with Midjourney (just with the text prompt) and got very different results. I'm not sure it's better/worse. Just interesting.

00400BB1-F0F2-49DC-9DBC-2E19AE068D32.jpeg
 
I find it really interesting how good the AIs are at creating superficially interesting art that fails completely at getting the details correct. Take that car above - looks great but then you realise that the front right wheel is all wrong (or perhaps non-Euclidean) - or those barbarians look fantastic, but every one of them has features that are bizarrely wrong such as the wonky spearhead or failed hands. All in a way that a human artist just wouldn't fail at. I wonder whether we'll see hybrid ways of working in future where the artist uses these tools to rough it out and then goes in to fix up these kind of errors and touch things up here and there?
 
I find it really interesting how good the AIs are at creating superficially interesting art that fails completely at getting the details correct. Take that car above - looks great but then you realise that the front right wheel is all wrong (or perhaps non-Euclidean) - or those barbarians look fantastic, but every one of them has features that are bizarrely wrong such as the wonky spearhead or failed hands. All in a way that a human artist just wouldn't fail at. I wonder whether we'll see hybrid ways of working in future where the artist uses these tools to rough it out and then goes in to fix up these kind of errors and touch things up here and there?
Hands are one of those things artists struggle with. Have you met a cartoonist? there's a reason most hands are either hidden behind objects or very simplified :smile:. I agree this tech could really be used to streamline illustration work.
 
I've had more fun (and success) with Midjourney. Here's a super team I am pulling together:

VERDIGRIS
Screen Shot 2022-09-02 at 10.23.09 AM.png

GUERILLA APE
Screen Shot 2022-09-02 at 10.22.34 AM.png

TANK CONCRETE
Picaroon_Jack_comic_book_art_of_patriotic_superhero_standing_in_d5e5bf72-83e1-449c-8fc0-0b288d...png

DOT COM
Picaroon_Jack_Comic_book_style_art_of_young_girl_hacker_smoking_1e6e79e0-ee8a-4776-91fc-e0e9fa...png

ACE OF SPADES
Screen Shot 2022-09-02 at 4.27.38 PM.png


Midjourney is not too keen on hands either. And I cannot get it to make sharks, dinosaurs, or airplanes. I got so frustrated with sharks that I just put in the prompt "shark" and here is what I got:
Picaroon_Jack_shark_efd0cb7f-1c39-4862-a0df-4f184da15123.png
 
Last edited:
I find it really interesting how good the AIs are at creating superficially interesting art that fails completely at getting the details correct. Take that car above - looks great but then you realise that the front right wheel is all wrong (or perhaps non-Euclidean) - or those barbarians look fantastic, but every one of them has features that are bizarrely wrong such as the wonky spearhead or failed hands. All in a way that a human artist just wouldn't fail at. I wonder whether we'll see hybrid ways of working in future where the artist uses these tools to rough it out and then goes in to fix up these kind of errors and touch things up here and there?

The future is now!



While its popularity has recently exploded, I've been working with Midjourney AI since the early beta, and I'm very experienced with it.

Midjourney is a remarkable AI which can turn text prompts into imagery.

While it's easy to pump out the endless stream of largely random images we're all seeing on our social media feeds, I'm using Midjourney in a much more directed, curated, and hopefully creative way.

It's not some cheap replacement for a human artist on this project - I've been really taken by the otherworldly nature of what Midjourney in its current form can produce, and it's been a perfect match for the slightly surreal feel of The Silver Road and Maskwitches. Maskwitches has been a truly engaging conversation with Midjourney.
 
Sheesh! I forgot about this one and I actually backed it.



In our last project, I experimented more with digital full-color art and I was happy with the evocative landscapes that came out of it. Continuing that progression, I've been exploring more digital tools like Midjourney. In conjunction with digital painting, the images produced give atmosphere and a tone, which is darker and different to the previous adventures.

Both of these Kickstarters are run by actual artists using Midjourney or other AI art to supplement their own art. I can see that being a huge timesaver if you can take a piece of AI-generated art that's 80-90% "done", import it into Photoshop and tweak till you're happy with it.
 
Hands are one of those things artists struggle with. Have you met a cartoonist? there's a reason most hands are either hidden behind objects or very simplified :smile:. I agree this tech could really be used to streamline illustration work.
Have to seen what some people do to feet? Paid professionals can't even do it. And muscular structure is hard too. See 90s comics for the struggle!
 
Some one just made this car on Midjourney:

markossi_a_Ford_hypercar_turns_into_translucent_cloth_blown_by__3b7bb392-60fa-4571-b19c-36b8e5...png

Here is their prompt:
a Ford hypercar turns into translucent cloth blown by the wind + round lights + extreme proportions + vibrant background + zoomed out + dynamic view + epic lighting + octane rendering + simple volumes + extremely detailed + ultra realistic + futuristic—ar 16:9 —v3 —testp —upbeta --upbeta​
 
I've been playing around with Dall-e for character portraits.
TQYHp29OW_dmUvWhs-rh-4hUGTSUkUjSkrNqWY9n1KiJq7PJuA3X5uv389wrMyclmPRxsQPgrnnMcQREpyDVNucOs4OiPDrYJ9LKSMYzWnvU6cN0rQtKoP3vTGAyzvJFFlRSIeZmr3rZ3KKPWCeI62gBdjr1hTKVVO4C6KPSC7BBq89VDeBc9rX_eloAv-Q9EM4GkzRscLYgR6NV6YDPmLqGKIS_ArUXDvlc2apHjsNNQOoMfGoVCcJBLxtxpTp81uxS34kumwSqDolTNe1u_n7HUmvoV9_myfDQ8L_NrqUbPz3c5o9HPH51PGcosj9sMkYBXH22WzUStEkpjLLUWOtu_V4_Ssv9V_RAjnYbqvxGzPqo9IwHvKxi-ptzD7yInWnM_yd6LwUblmRGLqwCURHHvr13YWlqlLlaC-sWem-i4jDIuF2hOWVLEEfwFBHE6tyMyjSuAz6jCWSNRQ_sNBmcFarr2vz3adF-GLzdWdgJtVS3KMD5OAB_KIid5hwFcx64inld7yv10TnbMxCwEcpyDDaGuG_ztrIpTie9OU2h2cBUzLTTyVDt-OaDMxNjh28dsv4QaerShAbyz40nTRpmDmgJE1Q6L5nSETHmSPGF3Tg9z4qt00u99JL19_9dze8CYNaK7ew9cZy4DtjyI5NOSWWd2fOpdN8-ngpFL_ggucBFBZsuT8Tte18Mpvs8cSMxNvvmwIjeA0UVTXV2RVKdpHHDFoX_iAKIoi56l3BDEMzI36JcE3zKmZWiUboqUF0MVnPlCT9w93h5z4JhWaUS381jeD0z-BLhVRakHmbsHQAq2LdhiyDDNqwB6E1NKV3PYg=s907-no


xUbxTeU1WD42ZLM_4elEt8LoGnBsKp6TeTjs0cOjyTSIFSDLlIjvZRg2NQ_4htxF54rv7dLh84yKTEFB1x269_7T5p9knCFk0_Q3Zs7A4oQ2rIdLjYi3EfAK8yz5Xt2dY648UGh3Pu0GPlFqwaT3b-t1_4zkkWp4eSrW98fRRIsmLaENn0C5wMOSVnn6BoCuKLkNPATkQQyWgjRJtJHjX7kB2wOe8k_vv705JEVg89aDY6yHP6xGWgIK5UWVBdmdqYTXFBgRMAcaljyoV1kQUxM1v_xvqLF3kM1N4COLpg7QAIZKN3gHzpT27zpHnUYUBAZuPV8bgjymwzYytHuj5I8P_81LSHkyLr5HG5VXurOlBD1mlobqSbnn8YjdLliWqe9kb53ep9ofVAKl7wuYluggEo8eLpRq442GXQeiQYaa6zom-ex3PMrTLZDJZjsOIJn6c1xPbc5srhcRyyrUPAZozjSwmPa8wSXXxyDuvBlx1zQIyFZ-L5SkxILchDSxkbZakE7zqzDCMHtOy8hFRIA2QbMGPNA_aaRKU_15c9_7vhQxQDURvU3nkfiPoaRvrWlBsiTVmh6OVtLZZJhL2HXhr7tbBbRgecnTHLZbFQrebS5SRbt0-yu6cZMzNbuM6YE3eeZ2Q7f_S2m3AWRSphjuj0pqSDnOlj7FSmyyrXs6tn76CEDtLWMCUbPZgVbR_qpjm0JgGoBVf2zazOknEmJIJvbzV_zeIX3OHzyTcktFb3jV-Y-ZDr-S-z_uGhPsK3j6-GokrnQKU2JBClXwN4d-BGAr4LVxl_SF84DNN6HSwioL29t-AxrfVzcIXGETJWhI9Q=s907-no


jW-u-Hkt_NrrVnlOpWTbt7dnR8Y61MY7u6qkj2arkxB6IVAlhTeEoM_dSyVldJXbgG3445wDsOhmwpHKKOo5BQqBj4pDH_8LQez1e9QOFCzcHzSTC_dmjdbyheAVof6uOQ80vR8lW4kBrkdRO6_UOLQwodtEY43hzwjVe-1axyMCsU3xJWvME3u0U4cXtjqEyktbFPYxV47Eqz_p9SuJhlVsmmfgZReqgqg7zMxRlSuQeSMN9-DQ_xJK7SvEefFF37p8WeKxOxUj80KlmVNczOZArR2QMfmuguIju2c4MiGR5yU_IAAalU0UAlVRZFUMFNgjAy3-WujqezPUcmWmHOA3fxKpssHaxsV2J4drTtX4H85lpjqwMfMbobcQkGN-B-cz78RFhGQExMn139a46sY6dJUr1Q2WyNu4OBaGowFwn2occxtSRR25n2bxu2BOqJKf0rBebWVTKVvJKEZGjD4pEerwi-R8hvLUgund1RFrPZ8Il9kFdyARY856nUlzAtyvkKxgVGe1JreP1pMUR0799FctnzquyA7HogtGO-1-ncpm3w-6uAnb6jhnuVWUOKk1sdCCOVQ-iLbL6YNh8OX8BE7pdIRSUJt4yEDPSveveQNpSWKTUw_bW2S2qMXmfl559OsfHWCAsnSj-7Gmh3BGqiIuiv842x4a7vsoJkwUDa2I1PzVRQAfrfMLeiRC1kz_5I3Wn6YTrFMRO_EZJJH6LdiU4pCqfvAlDDXo9feG82J1L3Hg5v3STnR88ZsD0hsQk9aZ2oagb7OgE7UIukA-EH0l7DHFuvYqkDztUsGLYhjGbUtRA1LcTFNdklT9LpGZPg=s907-no
 
DALL·E 2022-09-06 15.19.04 - psychic knights battling on a volcanic plain, illuminated by a gr...png
DALL·E 2022-09-06 15.15.50 - a surreal dark fantasy landscape, demonic, death knight, volcanic.png
DALL·E 2022-09-06 15.15.37 - a surreal dark fantasy landscape, demonic, death knight, volcanic.png

First one:
Psychic knights battling on a volcanic plain under a green sun: pretty goddamned close to what I was thinking about

The second two:
Surreal dark fantasy landscape, demonic . . . not exactly what I was thinking about, but still pretty cool nonetheless. Glad I found a way to quickly generate some nightmare fuel to help me envision some scenes or drop on my players when words just don't quite capture the essence of what's going on.
 
Welp, hopefully all those artists can cross-train as teachers or something. Mebbe they can still make a few bucks on Fiverr drawing customized anime porn (on that site recently and holy cow are there a LOT of ads for this). If these nascent tools can already produce pieces like I've seen here and elsewhere, basically for free or cheaps, I see no viable future for human professional artists.
 
Welp, hopefully all those artists can cross-train as teachers or something. Mebbe they can still make a few bucks on Fiverr drawing customized anime porn (on that site recently and holy cow are there a LOT of ads for this). If these nascent tools can already produce pieces like I've seen here and elsewhere, basically for free or cheaps, I see no viable future for human professional artists.
The limited number of folks this will put out of work were already not making a living doing this. It makes cool stuff for what it is but it's got limited ability to do exactly what I or someone else wants. This would have taken a job from the guy who illustrated OD&D. By the time AD&D was coming out TSR was making real money and the cost of an artist is not going to break a product.
 
Welp, hopefully all those artists can cross-train as teachers or something. Mebbe they can still make a few bucks on Fiverr drawing customized anime porn (on that site recently and holy cow are there a LOT of ads for this). If these nascent tools can already produce pieces like I've seen here and elsewhere, basically for free or cheaps, I see no viable future for human professional artists.
I do freelance editing work on RPGs, and I've had a few professional artists as friends over the years. That means I'm familiar with the back and forth process between an art director and artist, and I can't see many art directors being content with hoping the magic box spits out exactly what they need, not when they also have the option of paying a professional who reliable delivers on a deadline.

Also, to get the best results with these programs, you want to have a good knowledge of artistic styles and terminology, as well as a good grasp of ideas like color harmony and composition. You'll also need to know how to use photo-editing software to clean up what it gives you. In short, if a company decides to rely on a program to makes its art, it's going to need someone on staff to write the prompts, and edit the results, and the most qualified person is going to be an artist.

Artists survived the invention of photography, and they will survive this.

Dall-E now has allows you to upload images and expand and combine them. Here is my first, sloppy experiment to see what I could do with it.
Pea2lRNCDorNNnuJnWxq7LvGQ-qKamR1fVVKWS3bZ71bjEGMUJXw-anawTccQwNh6Oo1O-MlY7QY2lP8LnMzTjXA21U_cyASBx-773O2khCwyqwUb7VvN0pW5sIN_sTwr9Zx_79Hyycj2vN0inAYBBVeodFsQDEMbpaExaqB20pOaQAARckIcz09Lx0IyMJszWdoG-f1Nh4oBRGIB8R8DnYFGtrkEY3Ild-0Xop8pRdlb1aRpFsVib068kK48LJviZfnC3cJdHaQMhhc8iNh4bHdvNNK14ysqJOMdpF0Spj4x4TC1gnRDuWMDIG-NKCXcmJk-QXZ1XKYVBXjpHqwEMEuXDwFqrT510ukBFciIiXbbSnTV7kUOGoabqLfXdCWTl7yV8UoUwAcYSIqa_JMyG-mOxzXXgwKuHlQMkr8yohLadwb6lrU938mvubkktJiqfvIeCL2UMgWbR5odXqgdh5v2ZFNrzxzFeftJ8KPYs4FdJ3zt4xMiCQteo2TCzjQoH5sfwhblUDKaLd2Z0KdcUiORpzInyxmn6nrWCdG246yZYJponQaSiQzUfVn-QCJJND9eInGO3hF6Kz2P8GASme8WzjNj5zbYXr_nBmoYt0HWpff6tUTT1ai9soXQPJ2EY2Vt8OahrAPZwwrKsJ2hrah-ZTiqFWf5Pe-fkPioYsiVeO6jCUd8XKubyBoDwOfnbGgpHK-J0_Kqt183myN-7wq-2ddRnwtPnkXFacTHmTLj9y4XZmjLjApHFJM-LpTY-g_Ht5QhPcY7XS7eGf8uaSvAgOQr-uHaEM4Fcmi74JEJnDL3geXA_7jQWJAkddiCvkp2Q=w1707-h766-no


And here are a couple of somewhat more restrained follow-ups.
oNMXimxZgYyalaRiRZrQL3HVbkrFsprjVwHAaQRCcc1XCbjyzp3XSj3qcHxfBDQOhNnvP4IENmZ0m8GtlB5s8vB9hy87YhXRiBqiDEDq8c5wabV1pB6-eHiEAoZimrOzz7sWdGx3MQuAYcbX3eZ5iZqE1SLv2Ufmyhu_y8tZL1isns6vPoXJOZ4z2VGatSVHgmbPwGJZppuEGedRO9Dkn19iDgmhFi7egnVzSJBxC7U7IdcXM13CCsS4nsIjDYL3uudaHyB8Q0lPd16dDseooxRQ9Gyf98jNkM_XmcgQhabUsRH42d_Lvav17xxvPhhpVbsWi53lPpBPbBPb8ASgFOfRiy4Dcm4M63tuePcV2X32jzVyCN21Zaz-LPbvFkUaSfctzqDU1FcWM4hZYj4Y8d1to8EqoLPTh2jMTtquLaOYbMvT8OSI9WaEg3XaSOrj-m4znwLzdKRFZmQqsPVvjhsEHnNkUITDMvEtyeVT0eFQBieofdPRSL2RSdHWfazBN59sJUtik3v7AuCSg7a4cRw1vKIxex3Ig-RvsxAy1UL-navbENjUrss9PeuZnHJ_PUvXAzwGfSlNNJOy_rYfVziytzerxXLZkN2f5njPI5xoc6sCKXcP8wdGi4R8zsymP0SkHdrlO8muVDYT6OiCgYMxsTiXbIj-P__wAuSjpXVkDd1W3XLLvo6E_lKdJSivZMk8XZ4SIjJ69Kk3u_YMWUashLOp8Zqxs7gE0bNZ-qtA6HZJ15zeh83g8nPgkeuWThXFNoHjrEubgwxEhRcakv5YxTwYpUYZQ5amWw1HgME0tL15MzIzajPFp7eokUkR3QSaIw=w1248-h907-no


y82OOkP0L-pYn2KHOON_EGgRkjgqYP2cJWlgmtIXezwOXXfxwRjQ_grIfzTFRRJTgyQKoXtIaUK4d7jFDsPIuoyoHW7YQlWPd-x9HWjdN2GMAOfwrMcTCTEnQji80o21EPWy3EDLObGBUDdxGQJ-TTclULgYiRulv8BObWnSgyexxVvcNcEjp3bR9CZaQSqJgzsRbKktqNOPJ_X3AwtpBcLS76EEoawRKsuSuuvWXxmpPtlFP-q2daDVgwclzU5kxaV1h9CLqdQKyXFQeyioZY7G0tUHDBgX1j3Fjj5MsiflmETwS3lFEp2Xaw59fJIIm3cRhHn2TcLdm7lKPidMbEpPsv2jV87swCEH3dBO1HJhEXPaH9mRj6wrwJDX0MVIHqIFnR5NwvzW2P3YY1q_BT02a1eaTZY8JofzlAc4YUosm_BZXDjsP6RoNXG6zq6Np0HhppmbN4TnOlQhJau9kIfZERg66Bw_k_suaH9ufBYVSrEex2xMz91ZMTr3IKjr6irfWQ8auCGYX31phYGwxFD-0K_BRaIn8_FJHt_4kohL09unOiYKKUvMJFY31AE-5-fO2BYpj5nA1iAf18yi3u2qCh1khQY9JWHM2PoAr4siN1HUf0Jc4qjQ_F5Uy3-NStyOCylOz0t5mzFBrnvhI0rTw3nsX9TBfXqY0XjVBBpIqRinQFmWdPCoChNg15dEAeOHzUnf0jK8s2nsWKfbENNxZeMnUmxscymjcke4cH7Jf9L13VOc8IRkGdUz1M6pvW-AZ8c2QYA_5lF34vr-Dti6mfHevj5_M_4jkkSVji4aE17lcWeYSgbE2xUHzheJnayavA=w1503-h907-no


I'll add, it took a lot of prompting, rendering, erasing, re-prompting, rendering, and re-erasing to get these this point. Even with AI, there is still a lot of work involved.
 
Welp, hopefully all those artists can cross-train as teachers or something. Mebbe they can still make a few bucks on Fiverr drawing customized anime porn (on that site recently and holy cow are there a LOT of ads for this). If these nascent tools can already produce pieces like I've seen here and elsewhere, basically for free or cheaps, I see no viable future for human professional artists.
What I've seen leads me to the opposite conclusion.

All of the art lacks a level of precision and AI can't do context. That means it can't work accurately to a brief. And that isn't something I think is going to change in the foreseaable future. (If it does, we're talking flying cars level of advancement).

Largely, this is going to be used by people who would previously have used stock art or public domain works. It is a threat to some artists, but only those at the bottom of the professional food chain; the kind on Fiverr.

Something similar happened with copywriting with the introduction of unskilled workers who worked for peanuts into the equation. Right at the start, work dried up for people. But within a year clients realised that they just weren't able to get high quality work that way. And the most sensible move there for writers was simply to refuse to compete at all. If someone claimed they could get a 500 word article for a fiver, I wished them luck. (And if I got people asking me to clean up work that wasn't good enough because they'd tried to lowbudget it I charged them my full cost as if I'd done it from scratch. That has become less and less common).
 
What I've seen leads me to the opposite conclusion.

All of the art lacks a level of precision and AI can't do context. That means it can't work accurately to a brief. And that isn't something I think is going to change in the foreseaable future. (If it does, we're talking flying cars level of advancement).

Largely, this is going to be used by people who would previously have used stock art or public domain works. It is a threat to some artists, but only those at the bottom of the professional food chain; the kind on Fiverr.
Since art is subjective there is always going to be a use for AI art as generated, the problem is that AI is not intelligent. Since it started to become a thing in 1960s, various none of AI algorithms and techniques has produced anything that smacks human-level intelligence.

Instead, they are sophisticated tools.

For example, the expert system setups that originated in the 1980s are basically a series of algorithms and techniques that when applied allow a team to acquire some piece of expert knowledge and then structure it in a way that allows a tool to be made that a human can interact with in a more or less natural way to ask questions within that field and get an accurate answer.

It is not magic and it is not intelligent. It is just a setup that reflects a deep understanding of how people with expert knowledge organized and process information. The quality of the setup is utterly dependent on the team interviewing experts properly and setting up the initial data correctly.

The current round of AI Generation programs are in essence random tables that use advanced pattern recognition algorithms to tease out patterns that would be difficult for humans using manual methods to quantify involving visual images.

As a result, they are exactly as useful as an RPG random table would be. A good tool in many circumstances but when it comes to more sophisticated use like running a campaign or creating a product, you need a person that can use the results as a starting point for further work before it is up to snuff.

We already seen this with Traveller with any of the products detailing various sectors. The worst ones like Imperium Games, First Survey where the ones where the output just was accepted as is and then published. The original Spinward Marches suffered from this as well. The Solomani Rim was better because by then GDW learned that it was better to go over the results and massage the data to produce a more coherent product.
 
Another consideration with using AI is that you don't really know what sources the program is using and how close to the original it is. Remember when Exalted got busted using slightly altered video game art? It's a fair bet that people using these programs for commercial work are going to run into this problem.
 
Another consideration with using AI is that you don't really know what sources the program is using and how close to the original it is. Remember when Exalted got busted using slightly altered video game art? It's a fair bet that people using these programs for commercial work are going to run into this problem.
It is not a collage so the problem is exactly like if you asked an artist to paint a picture of a cat wearing sunglasses in the style of a starry night.

If the training data was limited then the number of patterns it can draw on is limited so the output looks sameish. But the practical reality is that to make these work even remotely well, you have thrown thousands of images at it. This is because no one artist can have enough drawings to train the algorithm well.

For example, the only reason it would know how to render a cat with sunglasses is that it had enough images with cats, with sunglass to infer enough pattern data to draw a cat or sunglasses in any style.

That is the problem was with this series of images based on a piece of art I bought. It had not context to know what the first image was about and just came up with some guesses based on proportion, line art style, and individual shape elements.

To do this right, there will be to be a generation of software that has to understand that it is a balding adult male wearing medieval clothing holding a scroll wearing a belt with a belted loop, pouch and a dagger.

They are probably working on that.

1662573265710.png
 
It is not a collage so the problem is exactly like if you asked an artist to paint a picture of a cat wearing sunglasses in the style of a starry night.

If the training data was limited then the number of patterns it can draw on is limited so the output looks sameish. But the practical reality is that to make these work even remotely well, you have thrown thousands of images at it. This is because no one artist can have enough drawings to train the algorithm well.

For example, the only reason it would know how to render a cat with sunglasses is that it had enough images with cats, with sunglass to infer enough pattern data to draw a cat or sunglasses in any style.

That is the problem was with this series of images based on a piece of art I bought. It had not context to know what the first image was about and just came up with some guesses based on proportion, line art style, and individual shape elements.

To do this right, there will be to be a generation of software that has to understand that it is a balding adult male wearing medieval clothing holding a scroll wearing a belt with a belted loop, pouch and a dagger.

They are probably working on that.

View attachment 49472
So I used to work on something like this for text. Our software used ontologies and taxonomies to understand the meaning of concepts and therefore could actually have some understanding of what a series of words in a sentence meant. Most of what you see even in very complex AI doesn't do that as I understand(and I've been out of it it's possible I'm entirely wrong here). Most AI uses co-occurrence to determine links between concepts. That makes it easy to build a model because you can just scrape data and build the cooccurance models. I'm sure it's much more complicated now with more meta data and chained cooccurance. What I think it still doesn't do that we did was have a walkable hierarchy of concepts and their directional weighting and lateral bonds. We did that pseudo automatically with human review and adjustment.
The net of it is that automated systems I think still don't understand the meaning of a word just it's proximity to other words. How that relates to images is the program may know a forearm is attached to a elbow and an elbow to an upper arm etc but it doesn't know humans only have two firearms and they shouldn't be attached to the same elbow.
 
Another consideration with using AI is that you don't really know what sources the program is using and how close to the original it is. Remember when Exalted got busted using slightly altered video game art? It's a fair bet that people using these programs for commercial work are going to run into this problem.


Looks like some of the sources being used for multiple AI art apps are unauthorized medical images that can be used to identify the subject. Glorious.
 

Looks like some of the sources being used for multiple AI art apps are unauthorized medical images that can be used to identify the subject. Glorious.
Of course, the real problem here is why those pictures were on the Internet to be scraped in the first place.

I'm not surprised to see Getty banning AI images over copyright concerns. It also makes sense that art-hosting sites are banning them. It's easy to see some users automatically generating images all day and indiscriminately posting them, regardless of quality. I could see them getting flooded with NFT-level crappy images.

My current prediction: Some companies will attempt using these to replace artists. Given that even the best images I have made using DALL-E and Stable Diffusion still require clean-up, they end up hiring artists to do post-production work. At some point, the artists get tired of removing surplus legs and straightening asymmetrical eyes. They just start making the art themselves and telling the bosses the AI did it.
 
My current prediction: Some companies will attempt using these to replace artists. Given that even the best images I have made using DALL-E and Stable Diffusion still require clean-up, they end up hiring artists to do post-production work. At some point, the artists get tired of removing surplus legs and straightening asymmetrical eyes. They just start making the art themselves and telling the bosses the AI did it.

Then people will be amazed at the marvelous AI which is just a person pretending to be an AI.

There was a Russian robot a few years ago that turned out to be a guy in a costume.
 
Banner: The best cosmic horror & Cthulhu Mythos @ DriveThruRPG.com
Back
Top