Thread: AI Art Generation |OT| Midjourney and beyond
Official Thread
TBF you do get a wide variety of models. There are specialist ones out there: -
If Gollum believed in Climate Change I assume?
Close.
The first two are "Greta Thunberg looking cute". The third one is "Greta Thunberg smiling".
A Ton Berge was AI extrapolating in any way may I think.
Mountain King
Member
- Platforms
I don't know dick about AI generation, but in some roundabout way this thread got me to a site called Lexica and...holy shit, I'm going to sit here all day.
Can someone please do a kick ass AI generated "R" for @Rock And Roll as a new avatar?
Something that is still rather simple so he doesn't get overwhelmed, just an "R", but at least a bit different than the no avatar pic?
Something that is still rather simple so he doesn't get overwhelmed, just an "R", but at least a bit different than the no avatar pic?
I don't know dick about AI generation, but in some roundabout way this thread got me to a site called Lexica and...holy shit, I'm going to sit here all day.
Yeah, Lexica runs its own Stable Diffusion-based models. It's pretty cool. A few other sites are out there that do the same or are in the process of being set up. Valuable if you lack the setup to run an install locally or don't want the hassle of setting up A1111, which does require a bit of heavy lifting as I posted about earlier in the thread. However local installation does mean you don't have to contend with censorship and there are a massive amount of freely available models to download, many of which produce really impressive imagery.
- Platforms
Great thread - there's some stunning generations in here. I was giving this a look over the weekend and only just managed to get my first half-decent result thanks to @Kadayi's excellent guide on Stable Diffusion and a prompt that I lifted from a few pages ago:
Great thread - there's some stunning generations in here. I was giving this a look over the weekend and only just managed to get my first half-decent result thanks to @Kadayi's excellent guide on Stable Diffusion and a prompt that I lifted from a few pages ago:
Thanks dude. Glad you appreciated it. As ever a lot keeps changing with the AI space. I've been so busy with other stuff I haven't done much myself of late, although I try to stay abreast of the news. Feel free to post some more.
News regarding a Stable Diffusion update
Free to test here
Free to test here
SDXL Release & Overview
Stability AI publicly released its latest image Ai model SDXL yesterday
A big step up from SD1.5 and the misfire of SD2.1 which never really hit the ground running
You can read the blurb about it here: -
stability.ai
However, the main thing to know is if you want to run it locally you are going to need a GPU with at least 8GB of RAM (preferably Nvidia though A1111 might work with AMD).
Some very basic test images are below: -
SDXL will run in Automatic1111, however, from my testing the performance isn't great presently in terms of render time and you have to jump through some hoops to get it to work well, however so if you do want to check it out then I suggest installing ComfyUI. As an Interface it can look a little bit daunting, however, it's basically an image-processing flow chart system and once you wrap your head around that aspect it makes sense.
I haven't yet had a chance to test out how SDXL performs with InvokeAI yet, but I suspect it lies somewhere in between A1111 & ComfyUI
Sebastian Kamph has a couple of Videos, the first detailing how to install ComfyUI and the second about installing SDXL and getting it up and running (there are a couple of 6GB models to download). Note you can easily link ComfyUI to your Automatic1111/InvokeAI models folders so you don't need to double up: -
If you want to stick with Automatic1111 then Olivio Sarikas has a video about how to get SDXL up and running in that: -
I dare say the performance issues of A1111 will get ironed out in short order, however, the main problem is that whereas with ComfyUI you can simply load up different workflows dependent on what you want to do, A1111 frontloads everything and as SDXL doesn't play nice if a whole raft of SD1.5 extensions are installed you have to disable most of them, then reload the UI which is obviously a huge pain in the ass. On the flip side, A1111 is still the king in terms of user-friendly extensions such as image browser, and extra network previews.
Note that no SD1.5 embeds, Hypernetworks, Lora, Lycoris work with SDXL as it is built on baseline images being 1024 x1024 whereas SD1.5 was built on 512x512. With that said you can output images at say 1024x768, but at least one dimension needs to be larger than 1024 otherwise you will run into distortion issues.
Don't expect any sexy times from the base model. It is going to be a while before people within the Image AI community have trained up custom models fully for that sort of thing, however with that said some early custom models are already up at CivitAI such as AI legend Lykons Dreamshaper SDXL alpha, as well as a first stab by him at an Anime Art SDXL model:
civitai.com
civitai.com
civitai.com
A big step up from SD1.5 and the misfire of SD2.1 which never really hit the ground running
You can read the blurb about it here: -
Announcing SDXL 1.0 — Stability AI
The Stability AI team is proud to release as an open model SDXL 1.0, the next iteration in the evolution of text-to-image generation models. Following the limited, research-only release of SDXL 0.9, the full version of SDXL has been improved to be the world's best open image generation model.
stability.ai
However, the main thing to know is if you want to run it locally you are going to need a GPU with at least 8GB of RAM (preferably Nvidia though A1111 might work with AMD).
Some very basic test images are below: -
SDXL will run in Automatic1111, however, from my testing the performance isn't great presently in terms of render time and you have to jump through some hoops to get it to work well, however so if you do want to check it out then I suggest installing ComfyUI. As an Interface it can look a little bit daunting, however, it's basically an image-processing flow chart system and once you wrap your head around that aspect it makes sense.
I haven't yet had a chance to test out how SDXL performs with InvokeAI yet, but I suspect it lies somewhere in between A1111 & ComfyUI
Sebastian Kamph has a couple of Videos, the first detailing how to install ComfyUI and the second about installing SDXL and getting it up and running (there are a couple of 6GB models to download). Note you can easily link ComfyUI to your Automatic1111/InvokeAI models folders so you don't need to double up: -
If you want to stick with Automatic1111 then Olivio Sarikas has a video about how to get SDXL up and running in that: -
I dare say the performance issues of A1111 will get ironed out in short order, however, the main problem is that whereas with ComfyUI you can simply load up different workflows dependent on what you want to do, A1111 frontloads everything and as SDXL doesn't play nice if a whole raft of SD1.5 extensions are installed you have to disable most of them, then reload the UI which is obviously a huge pain in the ass. On the flip side, A1111 is still the king in terms of user-friendly extensions such as image browser, and extra network previews.
Note that no SD1.5 embeds, Hypernetworks, Lora, Lycoris work with SDXL as it is built on baseline images being 1024 x1024 whereas SD1.5 was built on 512x512. With that said you can output images at say 1024x768, but at least one dimension needs to be larger than 1024 otherwise you will run into distortion issues.
Don't expect any sexy times from the base model. It is going to be a while before people within the Image AI community have trained up custom models fully for that sort of thing, however with that said some early custom models are already up at CivitAI such as AI legend Lykons Dreamshaper SDXL alpha, as well as a first stab by him at an Anime Art SDXL model:
DreamShaper XL - v2.1 Turbo DPM++ SDE | Stable Diffusion XL Checkpoint | Civitai
DreamShaper XL - Now Turbo! Also check out the 1.5 DreamShaper page Check the version description below (bottom right) for more info and add a ❤️ to...
Anime Art Diffusion XL - alpha3 | Stable Diffusion XL Checkpoint | Civitai
Anime Art Diffusion XL Check the version description below (bottom right) for more info and add a ❤️ to receive future updates. Do you like what I d...
MBBXL Ultimate - v1.0RC | Stable Diffusion XL Checkpoint | Civitai
The new version of MBBXL has been trained on >18000 training images in over 18000 steps. It's probably the most significant fine-tune of SDXL so...
Indepth with ComfyUI & InvokeAI 3.0
If you want to into ComfyUI in a bit more detail Scott Detweiler who now works for Stability AI has put out a couple of videos going into it in a bit more depth to show off all that the software can do:-
As well as a video about SDXL with some prompting advice given the Model is significantly more finetuned than SD1.5 or SD2.1
Also, Olivio put out a video recently covering the release of InvokeAI 3.0 which is a nice alternative to A1111
Personally, I think A1111 or InvokeAI are great starting points for image generation whilst ComfyUI is perhaps more for people who know exactly what they are doing or for advanced finetuning of an existing prompt. However they are ultimately all useful in their own way,
As well as a video about SDXL with some prompting advice given the Model is significantly more finetuned than SD1.5 or SD2.1
Also, Olivio put out a video recently covering the release of InvokeAI 3.0 which is a nice alternative to A1111
Personally, I think A1111 or InvokeAI are great starting points for image generation whilst ComfyUI is perhaps more for people who know exactly what they are doing or for advanced finetuning of an existing prompt. However they are ultimately all useful in their own way,
Better Photographic Results in Stable Diffusion
I must admit I tend to noodle more with the semi-real/stylised models than the photorealistic ones, however, I chanced across this video in my youtube feed today and the guy has some great tips for getting quality Photographic results with some of the better realism models like Absolure Reality and Photon when it comes to prompting: -
He also has a free (or pay what you want) PDF Guide up at Gumroad which is really good and well worth getting: -
These were just done using his prompt recommendations using the SD1.5 Model
His approach should be good for SDXL as well though obviously less the negative prompts as they aren't required in the new model. I'll endeavour to post up some tests over the weekend.
He also has a free (or pay what you want) PDF Guide up at Gumroad which is really good and well worth getting: -
These were just done using his prompt recommendations using the SD1.5 Model
His approach should be good for SDXL as well though obviously less the negative prompts as they aren't required in the new model. I'll endeavour to post up some tests over the weekend.
Put some time into SDXL using ComfyUI. Trying out some fashion magazine style prompts inspired by the photographic guide I posted about yesterday. Given SDXL is just a base model pretty impressed with the raw outputs which are pretty phenomenal (although it clearly struggles with the concept of DogTags as evinced below 
). Still, People within the SD community are already now releasing custom models based on SDXL on Civitai
). Still, People within the SD community are already now releasing custom models based on SDXL on Civitai
- Platforms
Good stuff again, @Kadayi - I'm going to have to spend a bit of time getting my head around ComfyUI and all the various nodes it provides I think. I did download someone's SDXL workload, but I'm not having much luck getting it to generate quality pictures so far without fully understanding the workflow for myself.
And using Automatic1111, I keep forgetting to switch from the refiner model and keep ending up with weird shit like this:
I did manage to get this from ComfyUI though after a bit of tweaking (even if the horse does have 5 legs):
And using Automatic1111, I keep forgetting to switch from the refiner model and keep ending up with weird shit like this:
I did manage to get this from ComfyUI though after a bit of tweaking (even if the horse does have 5 legs):
Just a quick heads up Civitai.com is running an image generation contest for SDXL the grand prize for which is a 4090 GPU
civitai.com
All you need to do is sign up to the site (It's free) and then generate some kick-ass SFW images and upload them to the contest page.
My suspicion is that StabilityAI themselves are sponsoring the competition to encourage people to give SDXL a whirl and come up with some high-end creative imagery to promote the new model, as a viable rival against Midjourney.
ENDED - SDXL image Contest! | Civitai
This Contest has now ended! SDXL Image Contest We've listened to your valuable feedback concerning the entry level for the SDXL training contest. I...
All you need to do is sign up to the site (It's free) and then generate some kick-ass SFW images and upload them to the contest page.
My suspicion is that StabilityAI themselves are sponsoring the competition to encourage people to give SDXL a whirl and come up with some high-end creative imagery to promote the new model, as a viable rival against Midjourney.
Just a quick heads up that Automatic1111 version 1.6 has dropped and with it some new features
Olivio talks about the early access, but the main version has just shipped today, so you can just update using Git Pull in SD install folder via the command prompt
Some cool features and improved UI.
The ability to use one model for the initial prompt and then another for the high res fix is kind of intriguing, plus there are a bevy of new samplers as well. Not tried out SDXL on it yet, though it's supposed to work better, however I would envisage it is still lagging behind ComfyUi in that regard, although ComfyUi obviously lacks some of A1111 nice features.
Olivio talks about the early access, but the main version has just shipped today, so you can just update using Git Pull in SD install folder via the command prompt
Some cool features and improved UI.
The ability to use one model for the initial prompt and then another for the high res fix is kind of intriguing, plus there are a bevy of new samplers as well. Not tried out SDXL on it yet, though it's supposed to work better, however I would envisage it is still lagging behind ComfyUi in that regard, although ComfyUi obviously lacks some of A1111 nice features.
Grisham
Ensuring Transparency
Grisham
Ensuring Transparency
Grisham
Ensuring Transparency
Grisham
Ensuring Transparency
Norman Reedus from the video game Death Stranding, octane render, photorealistic, high detail, unreal engine.
Grisham
Ensuring Transparency
- Platforms
I tried pixai yesterday and I think I need to save up for a PC that can comfortably run Stable Diffusion now. Wow. While pixai doesn't allow porn, they still let you create hot anime girls. I used my limited tries to make me some Kirigaya Suguha pics and the results were gorgeous Here's my favorite:
I mean, how amazing is that? That's a hotass, good quality drawing that didn't exist before I had it generated using my text prompts. Ai art changes everything
Here's my favorite:
I mean, how amazing is that? That's a hotass, good quality drawing that didn't exist before I had it generated using my text prompts. Ai art changes everything
Grisham
Ensuring Transparency
Grisham
Ensuring Transparency
I should really play around with it, but haven't had the time, but nice outputs.
