AI Image Generators

[HiDef]

This thread is for those who want to have full control over what they create using AI image generators and to hopefully gain insight from other users. I am not an expert so I can only offer insight from my own experience with such programs. I still don't like forums being flooded with images from generators but the programs can be useful in creating my own art, especially manips.

OK, this turned out to be disappointing for me because I have an Intel Mac Pro with dual AMD GPUs, which has plenty of power, RAM and VRAM to run machine learning apps. The problem is the model coding (Xcode on Macs). From what I've learned so far, the Mac versions of these apps are written for ARM architecture (M-series chips) and other versions are optimized for CUDA architecture (NVIDIA GPUs) in a Linux or Windows environment. It's a double whammy for me because I have neither. I may install Linux in a dual boot configuration in the future but for now I can only speak from a MacOS point of view. I haven't used Windows in 10 years.

The following is from my personal experience with MacOS running FREE AI image generators locally (offline only) and not WebUI cloud based generators which are subject to censorship.

There are a few reasons why my system is not compatible with AI image generators e.g. text to image, image to image, etc.

1. My system is high-end but does not have ARM architecture, it is a x64 Intel system.

2. NVIDIA GPUs have what's called, "tensor cores" which are designed to accelerate deep machine learning.

3. AMD GPUs do not have tensor cores.

Offline AI image generators for MacOS that I've tested are designed for M-series (Apple Silicon) chips but can be run on some non-silicon systems, but the caveat will be poor performance and much slower rendering times. So far all of the generators I've tested consumed 100% of VRAM (single GPU with 16GB VRAM) and crashed when using LLMs. So far none of them recognize multi-GPU systems and will only recognize one of the GPUs.

The following are the results from testing on my 2019 Mac Pro | 2TB Apple SSD (1.5TB free) | 3.2GHz 16-Core Intel Xeon | 96GB DDR4 | Dual Radeon RX 6900 XT (16GB x 2) | MacOS Sequoia 15.5.

During testing I tried out several models in each app before making my determination.

1. Draw Things AI: Fails to render anything except random noise, no image.
2. DiffusionBee: Fails to render anything except random noise, no image.
3. Easy Diffusion: Runs locally in Terminal using browser API, Poor performance because it runs on CPU and does not recognize GPU, Does not render prompts accurately, Some prompts are completely ignored resulting in a very random and distorted image, very slow. poor image quality, Text to image and image to image only, no video. Not worth using because it's not compatible with AMD GPUs in MacOS environment. The work around would be to install Linux or Windows.
4. Diffusers: Text to image only, Very fast, Non-verbose prompts generate a decent quality image in just a few seconds, Very limited models to choose from, some models ignore prompts resulting in random image noise, Eventually crashes because it does not release VRAM after each generation.
5. Stable Diffusion web UI: Runs locally in browser API, May need to install Windows with DirectML, Not tested yet.
6. ComfyUI: Not tested yet.
7. AUTOMATIC1111: Not supported on Intel Macs but user claims altering scripts enables it to run on CPU: https://github.com/AUTOMATIC1111/stable ... 1290150405

After further research, this is a work in progress: https://www.tomshardware.com/software/a ... her-things

[jca5646]

Have you tried posting this entire thread into ChatGPT? ChatGPT is what I used and worked with to set up my (non-Mac)! device. If not, copy and paste it all into a new ChatGPT chat and see what it comes back with. It may not solve or answer everything, but it might help provide potential solutions or point you to potential sources to help you rectify the issue(s).

[HiDef]

Have you tried posting this entire thread into ChatGPT? ChatGPT is what I used and worked with to set up my (non-Mac)! device. If not, copy and paste it all into a new ChatGPT chat and see what it comes back with. It may not solve or answer everything, but it might help provide potential solutions or point you to potential sources to help you rectify the issue(s).

Yes I have and it gives a generic response. It says all the programs I listed will work on Intel based Macs with AMD GPUs but it's not true on my system. I'm still researching this so I may find one that works well on my system?

[superheroinelinks]

ChatGPT works for me but it's hit or miss, the more detail provided the better the results. I didn't specify purple gloves and it assumed RED ?

[Mr. X]

If you can get automatic 1111 to work that's where you'll get your best results. Sadly it overheats my laptop so I have a dedicated rendering machine for it.

[HiDef]

ChatGPT works for me but it's hit or miss, the more detail provided the better the results. I didn't specify purple gloves and it assumed RED ?

ChatGPT is cloud based and subject to censorship. Some of the AI video shorts I've seen are too violent to have been created in a cloud based app.

[HiDef]

If you can get automatic 1111 to work that's where you'll get your best results. Sadly it overheats my laptop so I have a dedicated rendering machine for it.

Yes, I'm looking into that one but I won't be able to test it until later in the week, maybe this weekend?

[HiDef]

Updated OP after researching Automatic1111. No support for Intel Macs with AMD GPUs. Added workaround for those willing to use CPU for generation. For me it's a flat out no, I refuse to run software that is incompatible with my GPUs. My GPUs run other AI software without issue and has very fast rendering times. It would be like subscribing to a HSI service but opting to use a dialup connection.

[Mr. X]

Updated OP after researching Automatic1111. No support for Intel Macs with AMD GPUs. Added workaround for those willing to use CPU for generation. For me it's a flat out no, I refuse to run software that is incompatible with my GPUs. My GPUs run other AI software without issue and has very fast rendering times. It would be like subscribing to a HSI service but opting to use a dialup connection.

Yeah its a stickler though, you could get and MSI mini and run that however I don't believe A1111 will use the onboard intel card. One thing you could do if you have a thunderbolt port is to get an external drive mount and an external power supply and a geforce card. But I get you want to use your mac.

[HiDef]

Updated OP after researching Automatic1111. No support for Intel Macs with AMD GPUs. Added workaround for those willing to use CPU for generation. For me it's a flat out no, I refuse to run software that is incompatible with my GPUs. My GPUs run other AI software without issue and has very fast rendering times. It would be like subscribing to a HSI service but opting to use a dialup connection.

Yeah its a stickler though, you could get and MSI mini and run that however I don't believe A1111 will use the onboard intel card. One thing you could do if you have a thunderbolt port is to get an external drive mount and an external power supply and a geforce card. But I get you want to use your mac.

I have 2 other Mac Pro's, a 2009 and a 2010 and they're both heavily modded and nearly maxed out. I spent many hours, days, years, modding and upgrading them. I still use one them but I'm done modding them.
I purchased the 2019 because I wanted something that could do everything I needed and more natively without any modifications. After installing the dual RX 6900 XTs, I thought I was done upgrading/modding for a while. Since Apple dropped support for both Intel and AMD, that's as far as I can go with the GPUs. They still out perform even the M4's in certain instances. My GeekBench Metal scores are at the top of the heap: https://browser.geekbench.com/metal-benchmarks

It does do everything I need and more but I've hit a wall with AI generation. From my own research, it does not look very promising.

[Mr. X]

Updated OP after researching Automatic1111. No support for Intel Macs with AMD GPUs. Added workaround for those willing to use CPU for generation. For me it's a flat out no, I refuse to run software that is incompatible with my GPUs. My GPUs run other AI software without issue and has very fast rendering times. It would be like subscribing to a HSI service but opting to use a dialup connection.

Yeah its a stickler though, you could get and MSI mini and run that however I don't believe A1111 will use the onboard intel card. One thing you could do if you have a thunderbolt port is to get an external drive mount and an external power supply and a geforce card. But I get you want to use your mac.

I have 2 other Mac Pro's, a 2009 and a 2010 and they're both heavily modded and nearly maxed out. I spent many hours, days, years, modding and upgrading them. I still use one them but I'm done modding them.
I purchased the 2019 because I wanted something that could do everything I needed and more natively without any modifications. After installing the dual RX 6900 XTs, I thought I was done upgrading/modding for a while. Since Apple dropped support for both Intel and AMD, that's as far as I can go with the GPUs. They still out perform even the M4's in certain instances. My GeekBench Metal scores are at the top of the heap: https://browser.geekbench.com/metal-benchmarks

It does do everything I need and more but I've hit a wall with AI generation. From my own research, it does not look very promising.

Oh God I hate MAC. They're dropping open GL and now the products I service have to upgrade. Pain. And then some libraries can't meet their "Safe" standards anymore and the original maker is caput. Glad I only have a Mac Mini for testing. One product I work on dropped Mac support altogether and I believe Poser did the same.

Someone would have to be pretty driven to make a Vulcan version of automatic 1111 for Mac.

[HiDef]

Updated OP after researching Automatic1111. No support for Intel Macs with AMD GPUs. Added workaround for those willing to use CPU for generation. For me it's a flat out no, I refuse to run software that is incompatible with my GPUs. My GPUs run other AI software without issue and has very fast rendering times. It would be like subscribing to a HSI service but opting to use a dialup connection.

Yeah its a stickler though, you could get and MSI mini and run that however I don't believe A1111 will use the onboard intel card. One thing you could do if you have a thunderbolt port is to get an external drive mount and an external power supply and a geforce card. But I get you want to use your mac.

I have 2 other Mac Pro's, a 2009 and a 2010 and they're both heavily modded and nearly maxed out. I spent many hours, days, years, modding and upgrading them. I still use one them but I'm done modding them.
I purchased the 2019 because I wanted something that could do everything I needed and more natively without any modifications. After installing the dual RX 6900 XTs, I thought I was done upgrading/modding for a while. Since Apple dropped support for both Intel and AMD, that's as far as I can go with the GPUs. They still out perform even the M4's in certain instances. My GeekBench Metal scores are at the top of the heap: https://browser.geekbench.com/metal-benchmarks

It does do everything I need and more but I've hit a wall with AI generation. From my own research, it does not look very promising.

Oh God I hate MAC. They're dropping open GL and now the products I service have to upgrade. Pain. And then some libraries can't meet their "Safe" standards anymore and the original maker is caput. Glad I only have a Mac Mini for testing. One product I work on dropped Mac support altogether and I believe Poser did the same.

Someone would have to be pretty driven to make a Vulcan version of automatic 1111 for Mac.

It's a love/hate relationship for me. On one hand, I love the engineering that goes into their machines, the Mac Pro is a work of art. On the other hand, there's plenty of reasons to hate Apple too. Planned obsolescence is my #1 reason to hate Apple. They certainly did not want to make the same mistake they made with earlier Mac Pros, my MP 4,1 is 16 years old and is still viable today. I can't imagine being a poor sap who paid full price for a maxed out 2019 Intel MP at the time of their release only to find out the new ARM architecture was already in the works.

Looks like I'll be dual booting Linux or Windows in the near future. I have some software that is MacOS only so I have to keep it.

[superbia19872]

Have you tried using civitai? I also have an AMD processor. It's been my easiest route.


I trained this LoRA on kryptonite peril. It's a little wonky but it's my first one. I'm working on a second iteration right now, but the first attempt overfitted so I have to cull and retrain which will take a few days to save up buzz for.


So far, this works best with cyberrealistic and animagine base models, with detail tweaker xl set to 0.45.

https://civitai.com/models/1714206?mode ... Id=1939858

If you get anything good, please post it to the LoRA page as that will give me buzz at no cost to you that I can use to train the next iteration which will ideally be more versatile, more accurate, and will have an organized tag set to accompany it. Please for the love of God, do not try to use v1.01 or 1.02. They are BAD.

[HiDef]

Have you tried using civitai? I also have an AMD processor. It's been my easiest route.


I trained this LoRA on kryptonite peril. It's a little wonky but it's my first one. I'm working on a second iteration right now, but the first attempt overfitted so I have to cull and retrain which will take a few days to save up buzz for.


So far, this works best with cyberrealistic and animagine base models, with detail tweaker xl set to 0.45.

https://civitai.com/models/1714206?mode ... Id=1939858

If you get anything good, please post it to the LoRA page as that will give me buzz at no cost to you that I can use to train the next iteration which will ideally be more versatile, more accurate, and will have an organized tag set to accompany it. Please for the love of God, do not try to use v1.01 or 1.02. They are BAD.

It's not just about AMD GPUs, I also have an Intel Mac which is the reason why these apps are not compatible. If I were running Windows or Linux, I could probably get them to work with my AMD GPUs? Are you generating through an uncensored offline API or through a cloud based generator in your browser? What are your computer specs, e.g. Windows, Mac, etc., also what processor (CPU) do you have?

[Mr. X]

Have you tried using civitai? I also have an AMD processor. It's been my easiest route.


I trained this LoRA on kryptonite peril. It's a little wonky but it's my first one. I'm working on a second iteration right now, but the first attempt overfitted so I have to cull and retrain which will take a few days to save up buzz for.


So far, this works best with cyberrealistic and animagine base models, with detail tweaker xl set to 0.45.

https://civitai.com/models/1714206?mode ... Id=1939858

If you get anything good, please post it to the LoRA page as that will give me buzz at no cost to you that I can use to train the next iteration which will ideally be more versatile, more accurate, and will have an organized tag set to accompany it. Please for the love of God, do not try to use v1.01 or 1.02. They are BAD.

It's not just about AMD GPUs, I also have an Intel Mac which is the reason why these apps are not compatible. If I were running Windows or Linux, I could probably get them to work with my AMD GPUs? Are you generating through an uncensored offline API or through a cloud based generator in your browser? What are your computer specs, e.g. Windows, Mac, etc., also what processor (CPU) do you have?

Using civitai would be using and online resource and their servers. Same with their movie maker. No different than using Krea except Krea censors.

And civitai just uses automatic1111 on its servers. Basically you have to buy points.

[korda]

You'd still have to research how to run SDXL and Flux with an AMD GPU, but your 16GB VRAM card has enough memory to run well if you can get it to work. A better alternative than Civitai is to rent a GPU online through a service like Runpod. From my own experience, credits on Civitai run out very quickly. Renting or running locally are the best options. And be prepared to spend money on SSDs and HDDs for storage, Loras and checkpoints consume a lot of space.

[HiDef]

You're right Mr.X, it runs on their servers and processors so it doesn't matter what GPU you have or system for that matter. All you need is an internet connection.

Yep, credits run out really fast. I generated 4 images 4 times which used up all but 5 credits, you need 24 credits to run it one time. Most of the images generated were distorted and had extra limbs, broken limbs, etc. One thing to note is nudity is not censored on Civitai. 4 of the images had genital nudity but had 6 legs or the top half of the body was backwards with the genitals facing forward, so it's not a very good one. I tried a different model each time and they all generated distorted images. Reminded me of a car wreck I saw when I was a young lad, seeing body parts twisted and mangled and partially nude from the impact. You'd have to be into necro shit to find any of it appealing.

I don't need to worry about storage, I have over 20TB between my 2 networked machines.

I managed to get 1 decent image out 16.

YFHDHMTSWYJ773HFVFYYJ0J8V0.jpeg (157.93 KiB) Viewed 35778 times

[superbia19872]

Have you tried using civitai? I also have an AMD processor. It's been my easiest route.


I trained this LoRA on kryptonite peril. It's a little wonky but it's my first one. I'm working on a second iteration right now, but the first attempt overfitted so I have to cull and retrain which will take a few days to save up buzz for.


So far, this works best with cyberrealistic and animagine base models, with detail tweaker xl set to 0.45.

https://civitai.com/models/1714206?mode ... Id=1939858

If you get anything good, please post it to the LoRA page as that will give me buzz at no cost to you that I can use to train the next iteration which will ideally be more versatile, more accurate, and will have an organized tag set to accompany it. Please for the love of God, do not try to use v1.01 or 1.02. They are BAD.

It's not just about AMD GPUs, I also have an Intel Mac which is the reason why these apps are not compatible. If I were running Windows or Linux, I could probably get them to work with my AMD GPUs? Are you generating through an uncensored offline API or through a cloud based generator in your browser? What are your computer specs, e.g. Windows, Mac, etc., also what processor (CPU) do you have?

I'm doing everything on civitai because it's easy and free. I can farm a decent amount of buzz daily just doing they're thing
Most of my generations only cost 10 credits. I just don't want to deal with workarounds. I don't know a lot yet. But these are my specs.

Intel core i5- 10400 2.9GHz 2904 Mhz, 6 cores, 12 logical processors, radeon rx 580, 16gb ram. Windows 10 64 bit.

[superbia19872]

You're right Mr.X, it runs on their servers and processors so it doesn't matter what GPU you have or system for that matter. All you need is an internet connection.

Yep, credits run out really fast. I generated 4 images 4 times which used up all but 5 credits, you need 24 credits to run it one time. Most of the images generated were distorted and had extra limbs, broken limbs, etc. One thing to note is nudity is not censored on Civitai. 4 of the images had genital nudity but had 6 legs or the top half of the body was backwards with the genitals facing forward, so it's not a very good one. I tried a different model each time and they all generated distorted images. Reminded me of a car wreck I saw when I was a young lad, seeing body parts twisted and mangled and partially nude from the impact. You'd have to be into necro shit to find any of it appealing.

I don't need to worry about storage, I have over 20TB between my 2 networked machines.

I managed to get 1 decent image out 16.YFHDHMTSWYJ773HFVFYYJ0J8V0.jpeg

Hey, not terrible. The model can do better, though. What was your stack and parameters? 20 is high. I'm getting my best results for 8-10 credits per generation.

When you get more credits, you can remix this image.
https://civitai.com/images/87331222

Try using it as a starting point for your model stack and parameter and then just come up with your own prompt. I also find it helpful to give my prompt to chatgpt and ask it to refine it for the model I'm using. I will dig into the training data and get the tag set for you as well to help your prompts but as long as you're using the trigger "krypto exposure" the model will activate. For now i have to build a commander deck before game night tomorrow.

[HiDef]

You're right Mr.X, it runs on their servers and processors so it doesn't matter what GPU you have or system for that matter. All you need is an internet connection.

Yep, credits run out really fast. I generated 4 images 4 times which used up all but 5 credits, you need 24 credits to run it one time. Most of the images generated were distorted and had extra limbs, broken limbs, etc. One thing to note is nudity is not censored on Civitai. 4 of the images had genital nudity but had 6 legs or the top half of the body was backwards with the genitals facing forward, so it's not a very good one. I tried a different model each time and they all generated distorted images. Reminded me of a car wreck I saw when I was a young lad, seeing body parts twisted and mangled and partially nude from the impact. You'd have to be into necro shit to find any of it appealing.

I don't need to worry about storage, I have over 20TB between my 2 networked machines.

I managed to get 1 decent image out 16.YFHDHMTSWYJ773HFVFYYJ0J8V0.jpeg

Hey, not terrible. The model can do better, though. What was your stack and parameters? 20 is high. I'm getting my best results for 8-10 credits per generation.

When you get more credits, you can remix this image.
https://civitai.com/images/87331222

Try using it as a starting point for your model stack and parameter and then just come up with your own prompt. I also find it helpful to give my prompt to chatgpt and ask it to refine it for the model I'm using. I will dig into the training data and get the tag set for you as well to help your prompts but as long as you're using the trigger "krypto exposure" the model will activate. For now i have to build a commander deck before game night tomorrow.

I tried it again today. I used 25 credits at 5 credits per generation. I did not get a single decent image out of the 5 tries. Thanks but no thanks.

[HiDef]

I managed to get a few a good ones from SocialSight.AI before I was flagged for adult content.

a_photorealistic_image_of_batg (1).png (1.49 MiB) Viewed 35719 times

a_photorealistic_image_of_batg.png (1.49 MiB) Viewed 35719 times

photorealistic_batgirl_wear_pu.png (1.3 MiB) Viewed 35719 times

photorealistic_batgirl_wearing (3).png (1.32 MiB) Viewed 35719 times

a_photorealistic_image_of_batg (5).png (1.43 MiB) Viewed 35719 times

[Malice]

I managed to get a few a good ones from SocialSight.AI before I was flagged for adult content.

Those got flagged for adult content? Really!?

[HiDef]

I managed to get a few a good ones from SocialSight.AI before I was flagged for adult content.

Those got flagged for adult content? Really!?

Yep, after about 5 or 6 images it flagged me.

Here's one I saved from Civitai:

E5634NMW8HP9E09SPFB56WB4Y0.jpeg (194.78 KiB) Viewed 35699 times

The prompt was more verbose but to summarize it was supposed to be Catwoman sitting on Batgirl's face.

[Malice]

I managed to get a few a good ones from SocialSight.AI before I was flagged for adult content.

Those got flagged for adult content? Really!?

Yep, after about 5 or 6 images it flagged me.

Here's one I saved from Civitai:E5634NMW8HP9E09SPFB56WB4Y0.jpeg

The prompt was more verbose but to summarize it was supposed to be Catwoman sitting on Batgirl's face.

Well, what a shame. Especially when I see this pic...

Is there any platform out there that has less strict ToS? And of course quite qualitative in terms of image generation?

[superbia19872]

I managed to get a few a good ones from SocialSight.AI before I was flagged for adult content.
a_photorealistic_image_of_batg (1).png
a_photorealistic_image_of_batg.png
photorealistic_batgirl_wear_pu.png
photorealistic_batgirl_wearing (3).png
a_photorealistic_image_of_batg (5).png

Civitai is where it's at if you have to use a web based setup. You won't get shut down there. I got flagged once, they cleared it, and I've carried on ever since.

[HiDef]

I managed to get a few a good ones from SocialSight.AI before I was flagged for adult content.
a_photorealistic_image_of_batg (1).png
a_photorealistic_image_of_batg.png
photorealistic_batgirl_wear_pu.png
photorealistic_batgirl_wearing (3).png
a_photorealistic_image_of_batg (5).png

Civitai is where it's at if you have to use a web based setup. You won't get shut down there. I got flagged once, they cleared it, and I've carried on ever since.

What good is it if it does not understand prompts? All these apps claim that the paid version works better, but if they can't offer a free trail version that works, I have no reason to believe the paid version is any better. I gave up on Civitai after 20 or so tries that only produced 1 mediocre image that I posted above. Show us some images you made with Civitai with adult content.

[superbia19872]

I managed to get a few a good ones from SocialSight.AI before I was flagged for adult content.
a_photorealistic_image_of_batg (1).png
a_photorealistic_image_of_batg.png
photorealistic_batgirl_wear_pu.png
photorealistic_batgirl_wearing (3).png
a_photorealistic_image_of_batg (5).png

Civitai is where it's at if you have to use a web based setup. You won't get shut down there. I got flagged once, they cleared it, and I've carried on ever since.

What good is it if it does not understand prompts? All these apps claim that the paid version works better, but if they can't offer a free trail version that works, I have no reason to believe the paid version is any better. I gave up on Civitai after 20 or so tries that only produced 1 mediocre image that I posted above. Show us some images you made with Civitai with adult content.

I hear you, it can be difficult to work with. I usually describe what I want to chatgpt and have it refine my prompt for whatever tool I'm using.

All created on civitai using various LoRAs and base models. They're still not perfect, but they're something we can build on as this technology improves. I'm never charging anyone for it if they want to use it or improve on it. Posting one of your generations to the LoRA page does give me a small amount of on site currency, though, which goes back into refining the models. I'm currently working on a refined krypto model, but I'm having some trouble with overfitting so it's taking some time. It's hard to organize the data because I'm just sifting through imagery that's hot as fuck for hours at a time and I'm sure you know where I'm going with this.

[superbia19872]

I managed to get a few a good ones from SocialSight.AI before I was flagged for adult content.
a_photorealistic_image_of_batg (1).png
a_photorealistic_image_of_batg.png
photorealistic_batgirl_wear_pu.png
photorealistic_batgirl_wearing (3).png
a_photorealistic_image_of_batg (5).png

Civitai is where it's at if you have to use a web based setup. You won't get shut down there. I got flagged once, they cleared it, and I've carried on ever since.

What good is it if it does not understand prompts? All these apps claim that the paid version works better, but if they can't offer a free trail version that works, I have no reason to believe the paid version is any better. I gave up on Civitai after 20 or so tries that only produced 1 mediocre image that I posted above. Show us some images you made with Civitai with adult content.

I'm sorry, friend I could've sworn i pasted the link.

https://drive.google.com/drive/folders/ ... z-t1aNGwuC