A thread for anyone to share AI generated art from their own prompts.
-
For all readers and writers
After much ado, our new tagging system is here and being rolled out to the userbase. Read more here!
AI Generated Art Thread
- Thread starter PhoenixMercurous
- Start date
- Created at
- Index progress
- Ongoing
- Watchers
- 12
- Recent readers
- 0
- Threadmarks
- 2
-
Statistics (1 threadmarks, 190 words)
User Total PhoenixMercurous 1 Threadmarks Reader mode -
Loading…
Introduction
PhoenixMercurous
Soaring through the stars
- Location
- USA
- Pronouns
- He/Him
This thread is for sharing AI generated art and discussing how to use the various art generators that are available. If you want to discuss the technology behind the generators, I suggest you check out the AI Art thread over in Science & Technology. My goal for this thread is to have a place for people who dabble with AI art generators to share some of their results without needing to make their own dedicated threads. Anyone who wants to make a thread for their own AI art is welcome to do so.
If you want to have an index of the art you've shared, please make an index post of your own and poke me to get a link added to the "index of indices" and the post threadmarked.
Here's a list of currently available AI art generation tools or services that I'm aware of. It's not an exhaustive list, so feel free to poke me with information about any tools or services you'd like to have added to the list.
Lastly, be mindful of the site rules, particularly rule 6 and how it relates to images, and please read over the thread guidelines.
If you want to have an index of the art you've shared, please make an index post of your own and poke me to get a link added to the "index of indices" and the post threadmarked.
Here's a list of currently available AI art generation tools or services that I'm aware of. It's not an exhaustive list, so feel free to poke me with information about any tools or services you'd like to have added to the list.
- Stable diffusion by Huggingface: this is the basis for most of the currently available models, but not quite all of them. There's a web-based demo for testing prompts and learning how to use the tool, though not all the options are available. Instructions for running it locally are on the GitHub. Here's a getting started guide, and here's a redit post with some advice and tips.
- NovelAI: a web-based subscription service that offers an AI writing assistant and stable-diffusion based image generation. The art generation model also uses the Danbooru dataset for additional training, so it can use Danbooru tags as inputs in the prompt.
- Artbreeder: Offers image tweaking using an older model and now text based image generation. Also allow using collages of clip art in conjunction with text prompts. Credit to Silent Optics.
- DALLE-E-2: A different model that is a web service, that might require a monthly subscription. Credit to Silent Optics.
- Craiyon: Formerly DALL-E mini, this is a free web service that tends more towards surreal results. Credit to Silent Optics.
- Character.AI: From what I can tell, this is primarily an AI chatbot service by ex-Google engineers, it also has some image generation capabilities. It's free, so they're probably making money from selling users' data. Credit to Silent Optics.
Lastly, be mindful of the site rules, particularly rule 6 and how it relates to images, and please read over the thread guidelines.
- Don't share art other people generated without their permission.
- Don't share art created by using image2image on other's art without permission.
- Please credit the AI art generator used to produce the images.
- Please spoiler large series of images in a single post.
- If you want something added to the index of indices or list of art generators, let me know.
Last edited:
Phoenix Mercurous's Index
PhoenixMercurous
Soaring through the stars
- Location
- USA
- Pronouns
- He/Him
The index for all the art generated from my prompts.
Additionally, my standard unwanted content is: loli, shota, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, username, blurry. Most of the list is from Baughn, so I don't take credit for any improvements it makes to image generation.
Edit: Given the current controversy over AI generated art here on SV, I won't be posting any of my images here for the foreseeable future (not that I shared many anyways). Other people are free to share in the thread and @ me to get posts threadmarked, though I may be a bit slow to respond. I've been working in the office more lately, so tend to be on SV significantly less.
- Cat-dog-fox, Master Chief from Arknights, and a Kyuubi-Dragon
- The magic of the "backlighting" tag
- Alexander and Ariel Heaviside
- Examples from working on Ariel
- Another OC, Veronica Heaviside. Now with the entire prompt!
Additionally, my standard unwanted content is: loli, shota, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, username, blurry. Most of the list is from Baughn, so I don't take credit for any improvements it makes to image generation.
Edit: Given the current controversy over AI generated art here on SV, I won't be posting any of my images here for the foreseeable future (not that I shared many anyways). Other people are free to share in the thread and @ me to get posts threadmarked, though I may be a bit slow to respond. I've been working in the office more lately, so tend to be on SV significantly less.
Last edited:
PhoenixMercurous
Soaring through the stars
- Location
- USA
- Pronouns
- He/Him
And now for some actual artwork! Most of what I've made is from NovelAI, including all the images in this batch.
First off, I made a cat-dog-fox. Just to see if I could.
The prompt was: (cat | dog | fox). Model was the standard k_euler_ancestral, I think. For those unfamiliar with NovelAI, this syntax is used to mix prompts. I've heard that, at a technical level, it's averaging the losses from the prompts, but I'm not certain I'm remembering correctly.
Someone on a Discord server I'm part of mentioned that "Arknights" can lead to some interestingly styled results with NovelAI or other models trained with the Danbooru dataset, so here's Master Chief from Arknights:
Prompt was simply: Master Chief, arknights. I believe this is from before I started messing with settings other than the prompt, so everything else should be standard.
I'll close this first post out with one my very first image from the random madness generator that is variations on (kyuubi | dragon):
No, I'm not going to explain the story behind that prompt. It's honestly less interesting than the images the prompt can generate.
First off, I made a cat-dog-fox. Just to see if I could.
Someone on a Discord server I'm part of mentioned that "Arknights" can lead to some interestingly styled results with NovelAI or other models trained with the Danbooru dataset, so here's Master Chief from Arknights:
I'll close this first post out with one my very first image from the random madness generator that is variations on (kyuubi | dragon):
PhoenixMercurous
Soaring through the stars
- Location
- USA
- Pronouns
- He/Him
One of the downsides to NovelAI is that it tends to produce generic-ish anime girls, sometimes even from prompts that aren't directly asking for that. For example:
The prompt was a bit of a mess: ((kyuubi | dragon)|arknights | kyuubi | dragon), kyuubi, dragon, arknights. This isn't directly asking for an anime girl, but it still produced one. Though I kind of like the mischievous look on the girl's face.
The unwanted content was probably my standard of: loli, shota, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, username, blurry.
I've found some tricks to help push NovelAI out of it's standard style, though they don't always work. The ones that work most reliably seem to be adding "backlighting" and "nose, lips" to the prompt, assuming you're asking for artwork of a person. For example:
Prompt was: 1girl, full body, face, lips, nose, eyes, backlighting, high quality, small nose, brown hair, long hair, curly hair, blue eyes. Unwanted content was my standard.
This is loosely based on a family member of mine, though I asked for a different eye color.
The prompt for this one was too long for a file name, so I don't have all of it. What I know for certain is: springfield (girls' frontline), green eyes, brown hair, long hair, nose, lips, collarbone, eyelashes, eyebrows, smile, cowboy shot, blue dress, indoors, curtains, backlighting,
I'm guessing it also included: photorealistic, hand,
Unwanted content was my standard.
I've found that "hand" usually works better than "hands". The later usually results in too many hands, while the former helps encourage NovelAI to do slightly better looking hands, though it rarely gets them looking terribly good.
The unwanted content was probably my standard of: loli, shota, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, username, blurry.
I've found some tricks to help push NovelAI out of it's standard style, though they don't always work. The ones that work most reliably seem to be adding "backlighting" and "nose, lips" to the prompt, assuming you're asking for artwork of a person. For example:
Prompt was: 1girl, full body, face, lips, nose, eyes, backlighting, high quality, small nose, brown hair, long hair, curly hair, blue eyes. Unwanted content was my standard.
This is loosely based on a family member of mine, though I asked for a different eye color.
The prompt for this one was too long for a file name, so I don't have all of it. What I know for certain is: springfield (girls' frontline), green eyes, brown hair, long hair, nose, lips, collarbone, eyelashes, eyebrows, smile, cowboy shot, blue dress, indoors, curtains, backlighting,
I'm guessing it also included: photorealistic, hand,
Unwanted content was my standard.
I've found that "hand" usually works better than "hands". The later usually results in too many hands, while the former helps encourage NovelAI to do slightly better looking hands, though it rarely gets them looking terribly good.
Last edited:
PhoenixMercurous
Soaring through the stars
- Location
- USA
- Pronouns
- He/Him
To wrap things up for today, here are some original characters I had NovelAI make artwork of. Both of the characters are from a collection of stories I'm probably never going to write.
First up is Alexander Heaviside:
The prompt was too long for a file name, so I don't have all of it. I know some of the key bits were: {{{skinny}}}, mature male, old man, tweed suit, very short hair. It took a bit of tuning to get the power of "mature male" and "old man" correct, but unfortunately those were in the part of the prompt that got cut off. Unwanted content was my standard.
In my very silly collection of story ideas, Alexander is the chief physicist of the group of protagonists and eventually the patron of atheists and, despite his claims otherwise, the god of time, death, and theoretical physics.
Next up is his youngest sibling, Ariel Heaviside:
This is yet another mostly lost prompt. I know it included "dynamic pose" and I'm fairly certain it included "yoh yoshinari" and "arknights." I forget if this exact image included the phrase "tactical blazer," but I know I tried that at some point. Unfortunately, Ariel's combat outfit is too complex for NovelAI to get quite correct, this is as close as I managed to get. I still like the pose, it does a pretty good job communicating her/their character.
Ariel's title is "Defender of the Innocent", a mantle she took up after a bit of a rough time as an older teenager and young adult. An Eva/Gundam fusion story will do that to characters. They're also the god of math and magic, though neither is showing in this picture. I've struggled to get good looking magic auras out of stable diffusion based image generation.
First up is Alexander Heaviside:
The prompt was too long for a file name, so I don't have all of it. I know some of the key bits were: {{{skinny}}}, mature male, old man, tweed suit, very short hair. It took a bit of tuning to get the power of "mature male" and "old man" correct, but unfortunately those were in the part of the prompt that got cut off. Unwanted content was my standard.
In my very silly collection of story ideas, Alexander is the chief physicist of the group of protagonists and eventually the patron of atheists and, despite his claims otherwise, the god of time, death, and theoretical physics.
Next up is his youngest sibling, Ariel Heaviside:
This is yet another mostly lost prompt. I know it included "dynamic pose" and I'm fairly certain it included "yoh yoshinari" and "arknights." I forget if this exact image included the phrase "tactical blazer," but I know I tried that at some point. Unfortunately, Ariel's combat outfit is too complex for NovelAI to get quite correct, this is as close as I managed to get. I still like the pose, it does a pretty good job communicating her/their character.
Ariel's title is "Defender of the Innocent", a mantle she took up after a bit of a rough time as an older teenager and young adult. An Eva/Gundam fusion story will do that to characters. They're also the god of math and magic, though neither is showing in this picture. I've struggled to get good looking magic auras out of stable diffusion based image generation.
Could you add sharing the prompts, settings of the generator etc as mandatory in the rules? @PhoenixMercurous
PhoenixMercurous
Soaring through the stars
- Location
- USA
- Pronouns
- He/Him
I've gone back and forth on whether I want to encourage people to do that or not. I left it out of the guidelines because I'm still uncertain.
I won't go so far as to say it's mandatory because it would be quite a chore to track everything for NovelAI. The only things it puts in the file names are the prompt (up to a certain character limit) and the seed. I'd have to make a spreadsheet to track the full prompt and all the settings.
I get why this works, but 'vanquish the anime by asking for a nose' still sounds like a shitpost.
That is a really nice pose. Whatever version of SD artbreeder uses would definitely not put a sword that is held outside of the frame at a reasonable angle.
For my pictures I will do a hubris/lazy and just assume the website is still by the time you are reading this by just pasting a link.
Pictures were selected by randomly scrolling through the 'created+liked' of my profile and picking stuff.
The AI knows what selfies are
It's so cursed that first-person descriptions are effective
Collages often end up great
Random words=Cool monster
PhoenixMercurous
Soaring through the stars
- Location
- USA
- Pronouns
- He/Him
It looks like Artbreeder has a rather different default style to NovelAI and whichever base version of SD I last played with.
With anything I post, it's important to remember that I'm only sharing what I feel are the very best results. For example, that one image of Alexander is the last from a chain of 50 images. It looks like I made 98 images of Ariel in that sitting, which was after I'd already worked out a prompt that generally got her/their look correct. I don't have all the images from the process of refining the prompt, but I'd guess there were somewhere around 100 images from that. Here's a smattering of images from the process of refining the prompt for that combat pose of Ariel, all from NovelAI:
The first chunk of the prompt appears to have been the same for all these images: amazing lighting, beautiful, incredible details, absurdres, dynamic pose, sword, short sword, holding sword, 1girl, woman, adult, solo, original, blue eyes, red hair, short hair, small breasts. Unfortunately the rest of the prompt, where I was making changes, didn't make it into the file names. I think most or all of the prompts included: yoh yoshinari, arknights, neon genesis evangelion, armor, breastplate, armor under clothing, jacket. I played around with the color of gloves in the prompt and tried things like "military jacket", "military uniform", and "tactical clothing".
Edit: I think this might've been before I snagged the big list of unwanted content, so it's possible the base unwanted content was just: loli, shota.
Last edited:
I guess I can share some more stuff so this doesn't die.
I love that right hand
Badly drawn hands can be very effective in general
I like how the AI does phone UIs
Background unfortunately doesn't mesh well, but I love the look of the figure
Clashing weird materials are an easy path to cool shit
I love that right hand
Badly drawn hands can be very effective in general
I like how the AI does phone UIs
Background unfortunately doesn't mesh well, but I love the look of the figure
Clashing weird materials are an easy path to cool shit
PhoenixMercurous
Soaring through the stars
- Location
- USA
- Pronouns
- He/Him
I worked a bit more on some of my OC's, and this time I copied the entire prompt down. The character is Veronica Heaviside, Alexander's wife. She's the team's best martial artist, expert on computer hardware, and god of hope. She's called god, not goddess, due to the quirks of her (and Alexander's and Ariel's) first language, which is an in-universe conlang. Maybe I'll try to use NovelAI to help me write the story some day.
Prompt: 1girl, (toned:1.05), dark skin, very dark skin, black skin, medium breasts, brown hair, medium hair, brown eyes, nose, lips, yoh yoshinari, greg rutkowski, (arknights:0.95), diffuse lighting, (muscular:0.9), full body, solo, purple shirt, pants, black pants, realistic, (neon genesis evangelion:0.95), dynamic pose, black footwear, fighting stance, spear, lance, buttons,
Unwanted content: tight shirt, tight pants, skin tight, cleavage, loli, shota, bad anatomy, missing limb, extra digits, worst quality, low quality, text, watermark
This might've been using the curated anime dataset, I'd have to rerun the prompt & seed to be certain. Scale, steps, and model were all left at NAI's defaults.
The departure from my longer list of unwanted content is because I'm using a different computer while traveling for the holidays. The additions at the start are to suppress NovelAI's annoying default of making characters sexy unless you say otherwise.
I'm liking this style (and the fact I wrote down the entire prompt), so I'll probably have NAI make new art for Ariel and Alexander in this style (or as close as I can keep it) at some point.
Prompt: 1girl, (toned:1.05), dark skin, very dark skin, black skin, medium breasts, brown hair, medium hair, brown eyes, nose, lips, yoh yoshinari, greg rutkowski, (arknights:0.95), diffuse lighting, (muscular:0.9), full body, solo, purple shirt, pants, black pants, realistic, (neon genesis evangelion:0.95), dynamic pose, black footwear, fighting stance, spear, lance, buttons,
Unwanted content: tight shirt, tight pants, skin tight, cleavage, loli, shota, bad anatomy, missing limb, extra digits, worst quality, low quality, text, watermark
This might've been using the curated anime dataset, I'd have to rerun the prompt & seed to be certain. Scale, steps, and model were all left at NAI's defaults.
The departure from my longer list of unwanted content is because I'm using a different computer while traveling for the holidays. The additions at the start are to suppress NovelAI's annoying default of making characters sexy unless you say otherwise.
I'm liking this style (and the fact I wrote down the entire prompt), so I'll probably have NAI make new art for Ariel and Alexander in this style (or as close as I can keep it) at some point.
- Location
- USA
View: https://imgur.com/a/8pUP8UY
I made this short story, I am not sure what it is about tho.
I used Stable Diffusion with InvokeAI+MS Paint+ Paint.NET
PhoenixMercurous
Soaring through the stars
- Location
- USA
- Pronouns
- He/Him
NovelAI seems to have updated their model, I think it's gotten slightly better at doing hands. It's still quite bad with weapons, though.
I had it generate a new batch of images of Ariel using the same style tags I used for Veronica, here's my favorite of the bunch:
The broad strokes are all the same, but the coloration is a lot flatter, especially on Ariel's face.
I had it generate a new batch of images of Ariel using the same style tags I used for Veronica, here's my favorite of the bunch:
I've found lowering the scale can help with keeping the model from going into it's "generic Danbooru not-quite-hentai" style. For comparison, here's what putting the scale back to the default values of 11 does when holding everything else fixed:
- Location
- USA
I am bit bored so here are some highlights from the last month.
Eery random shape
Really reminds me of the jaw worm from sts
Something less weird and just pretty
It's fun when stuff almost gets NSFW
Surreal floaty stuff
Eery random shape
Really reminds me of the jaw worm from sts
Something less weird and just pretty
It's fun when stuff almost gets NSFW
Surreal floaty stuff