Below is my general advice for getting good, uncensored txt2img AI results WITHOUT the need for an online AI service. Using the app Automatic1111 or ComfyUI, you can create unlimited AI artworks for free, if you have a PC with an NVidia 3000 or 4000 series graphics card.
My Base Setup
How to Source AI Info
Google, YouTube and Reddit are INVALUABLE sources of information when it comes to learning generative AI. I learned everything I know by performing searches on those platforms and reading/watching my way through the treasure trove of guides created by generous experts. I consider myself to be a fairly advanced AI creator now, and I did not have to ask for help simply because the information was so easy to find.
About Stable Diffusion
SD can be used online or you can install and run it on Windows if you have powerful enough hardware. I run it directly from my PC and consider that to be the ideal approach, since it enables me to fully customize my software quickly and without restrictions. Running locally also enables you create uncensored AI renders.
About Automatic1111 and ComfyUI
Stable Diffusion is an AI rendering engine. The two most popular and powerful options for using SD are Automatic1111 and ComfyUI, which are front-end apps for the SD engine. For beginners, I recommend Automatic1111. Right now there's a lot of buzz surrounding ComfyUI, but I believe it's completely over-hyped vs A1111. I have mastered ComfyUI and consider myself to be an advanced user, and I've come to the conclusion that it's overly complicated and cumbersome to navigate for what you get in return. A1111 can do pretty much everything that ComfyUI can do, and it enables me to work MUCH faster due to the optimized workflows and built-in drag-and-drop toolsets. I use both, but A1111 simply gets out of my way and let's me work quickly during multi-pass workflows. Note: This blog post was written in January 2024, and I expect that ComfyUI will likely add quality-of-life improvements over time.
Installing A1111 or ComfyUI
Installing either front end on your local machine can be a complex task, depending on your technical experience. There are MANY step-by-step installation guides on YouTube and I recommend you start there. Doing so involves using the Windows command line interface. If you lack command line experience, you can definitely follow a step-by-step guide to figure it out - but it will take some diligence and patience.
How to source SD Checkpoints/Models and LoRAs
I use SD 1.5 checkpoints and LoRAs. You can download several excellent checkpoints and LoRAs via the CivitAI website for free. Every checkpoint that I use came from CivitAI. I use their advanced search feature to filter results depending on what I'm looking for. They have checkpoints for various artistic styles and content types, including niche erotic content.
About training your own custom LoRAs
LoRA's are NOT necessary and can very easily dilute the overall quality of your checkpoint's output. I recommend avoiding LoRAs until you've gained some experience with your favorite checkpoints. That said, nothing beats creating your own LoRAs. I have created my own using the "Kohya SS" app. It is a complex process that requires fairly powerful PC hardware, but it's worthwhile if you're serious about stepping up your game and customizing your gens.
About My Basic Workflow
I employ a multi-pass workflow, using inpainting extensively. It goes something like this:
Hi-res fix is essential when using SD 1.5 checkpoints. Be sure to "activate" hi-res fix for img2img by going to the Upscaling section within A1111's settings and setting "Upscaler for img2img" to "SwinIR_4x". This enables you to apply hi-res fix via txt2img OR downstream via img2img. Depending on the composition of my image, I might apply hi-res fix early in my workflow using the txt2img tab, in which case I perform inpainting on the upscaled image. Conversely, sometimes I'll do all my inpainting on the low-res image, and then run hi-res fix on the corrected image using the img2img tab. Sometimes I inpaint the low-res AND the high-res image.
About ControlNet
Once you've got the basics down, watch ControlNet tutorials on YouTube. ControlNet is a powerful collection of tools that will let you control your images in a wide variety of ways. It is essential to advanced use of SD (whether you use A1111 or ComfyUI).
My Essential A1111 plugins and ControlNet processors
- Inpaint Anything + Segment Anything (for automatically generating inpaint masks)
- ReActor (for face swapping)
- ADetailer (for cleaning up faces if I'm not using face-swap/ReActor)
- ControlNet IP Adapter (for copying specific artistic styles and faces from existing images)
- ControlNet DWPose (for copying poses from existing images)
- ControlNet DepthAnything (for copying compositions/layouts from existing images)
- ControlNet Depth Hand Fixer (aka Mesh Graphormer - for fixing wonky hands)
Parting Words
If you're new to AI image gen, the above info should be enough to fuel a thousand Google, YouTube and Reddit searches, leading you to articles and tutorials that walk you through all of these techniques and more. I work in a creative field, so it has been worth my time to learn this stuff, but even if I didn't need it for professional use, I would have still learned it - because it has been a total blast. Note: I have a background in software engineering plus a very powerful hand-built computer, so getting everything installed and running was fairly easy for me, but I believe it can be done even if you're a novice. Your mileage may vary. Good luck, and have fun!
My Base Setup
- Hardware: PC - Ryzen 3950x CPU - 64 GB RAM - RTX 4090 GPU w/24 GB VRAM
- Software: Automatic1111 / Stable Diffusion
How to Source AI Info
Google, YouTube and Reddit are INVALUABLE sources of information when it comes to learning generative AI. I learned everything I know by performing searches on those platforms and reading/watching my way through the treasure trove of guides created by generous experts. I consider myself to be a fairly advanced AI creator now, and I did not have to ask for help simply because the information was so easy to find.
About Stable Diffusion
SD can be used online or you can install and run it on Windows if you have powerful enough hardware. I run it directly from my PC and consider that to be the ideal approach, since it enables me to fully customize my software quickly and without restrictions. Running locally also enables you create uncensored AI renders.
About Automatic1111 and ComfyUI
Stable Diffusion is an AI rendering engine. The two most popular and powerful options for using SD are Automatic1111 and ComfyUI, which are front-end apps for the SD engine. For beginners, I recommend Automatic1111. Right now there's a lot of buzz surrounding ComfyUI, but I believe it's completely over-hyped vs A1111. I have mastered ComfyUI and consider myself to be an advanced user, and I've come to the conclusion that it's overly complicated and cumbersome to navigate for what you get in return. A1111 can do pretty much everything that ComfyUI can do, and it enables me to work MUCH faster due to the optimized workflows and built-in drag-and-drop toolsets. I use both, but A1111 simply gets out of my way and let's me work quickly during multi-pass workflows. Note: This blog post was written in January 2024, and I expect that ComfyUI will likely add quality-of-life improvements over time.
Installing A1111 or ComfyUI
Installing either front end on your local machine can be a complex task, depending on your technical experience. There are MANY step-by-step installation guides on YouTube and I recommend you start there. Doing so involves using the Windows command line interface. If you lack command line experience, you can definitely follow a step-by-step guide to figure it out - but it will take some diligence and patience.
How to source SD Checkpoints/Models and LoRAs
I use SD 1.5 checkpoints and LoRAs. You can download several excellent checkpoints and LoRAs via the CivitAI website for free. Every checkpoint that I use came from CivitAI. I use their advanced search feature to filter results depending on what I'm looking for. They have checkpoints for various artistic styles and content types, including niche erotic content.
About training your own custom LoRAs
LoRA's are NOT necessary and can very easily dilute the overall quality of your checkpoint's output. I recommend avoiding LoRAs until you've gained some experience with your favorite checkpoints. That said, nothing beats creating your own LoRAs. I have created my own using the "Kohya SS" app. It is a complex process that requires fairly powerful PC hardware, but it's worthwhile if you're serious about stepping up your game and customizing your gens.
About My Basic Workflow
I employ a multi-pass workflow, using inpainting extensively. It goes something like this:
- I generate batches of 16 images at a time, experimenting with my base prompt until I start getting images that come close to what I want.
- I continue generating batches until I find that "needle in the haystack" image. I look for 3 things: Images with good composition, non-deformed hands and (if creating erotic content) non-deformed genitalia. These things can be fixed using inpainting, but the less inpainting, the better, since it can be incredibly time-consuming.
- Inpaint via img2img to optimize the face, hands, and any other details that need enhancing. Inpaint each area separately and tailor your prompt as you go. Most of my gens have gone through several inpainting passes. I'll run each inpaint mutltiple times using random seeds until I get what I want.
- Upscale using hi-res fix.
Hi-res fix is essential when using SD 1.5 checkpoints. Be sure to "activate" hi-res fix for img2img by going to the Upscaling section within A1111's settings and setting "Upscaler for img2img" to "SwinIR_4x". This enables you to apply hi-res fix via txt2img OR downstream via img2img. Depending on the composition of my image, I might apply hi-res fix early in my workflow using the txt2img tab, in which case I perform inpainting on the upscaled image. Conversely, sometimes I'll do all my inpainting on the low-res image, and then run hi-res fix on the corrected image using the img2img tab. Sometimes I inpaint the low-res AND the high-res image.
About ControlNet
Once you've got the basics down, watch ControlNet tutorials on YouTube. ControlNet is a powerful collection of tools that will let you control your images in a wide variety of ways. It is essential to advanced use of SD (whether you use A1111 or ComfyUI).
My Essential A1111 plugins and ControlNet processors
- Inpaint Anything + Segment Anything (for automatically generating inpaint masks)
- ReActor (for face swapping)
- ADetailer (for cleaning up faces if I'm not using face-swap/ReActor)
- ControlNet IP Adapter (for copying specific artistic styles and faces from existing images)
- ControlNet DWPose (for copying poses from existing images)
- ControlNet DepthAnything (for copying compositions/layouts from existing images)
- ControlNet Depth Hand Fixer (aka Mesh Graphormer - for fixing wonky hands)
Parting Words
If you're new to AI image gen, the above info should be enough to fuel a thousand Google, YouTube and Reddit searches, leading you to articles and tutorials that walk you through all of these techniques and more. I work in a creative field, so it has been worth my time to learn this stuff, but even if I didn't need it for professional use, I would have still learned it - because it has been a total blast. Note: I have a background in software engineering plus a very powerful hand-built computer, so getting everything installed and running was fairly easy for me, but I believe it can be done even if you're a novice. Your mileage may vary. Good luck, and have fun!
