HOME | DD

Description
How to create AI Art like me

1 Set up InvokeAI and run it.

If you're comfortable with getting python projects from github and building environments with conda or pip in virtualenv, you can do it yourself. It's not ready for mainstream otherwise. Check out invoke-ai.github.io/InvokeAI/ and follow their instructions to set up.

2 Get some ideas.

Sometimes I bring my own but I like variety and am creating cheap, fast, ephemeral art for my own enjoyment so I generate ideas and pick what I like. I have used various art prompt sites and made my own roll tables. Today I'm enjoying using a ChatGPT session primed to suggest art prompts. I primed the ChatGPT session with "I want you to act as an AI artists muse. You will help artists create artwork using by suggesting creative and interesting image prompt ideas for according to whatever theme I ask for. My first request is "I need prompts that feature mushroom girls having fun.""  Some back and forth later, ChatGPT and I narrowed what I wanted to create, and I selected the prompt "A serene and tranquil florist, with short, blonde hair and piercing, blue eyes, arranges a bouquet of beautiful and fragrant flowers, her movements slow and graceful as she places each blossom with care and attention to detail, her peaceful and nurturing demeanor calming."

3 Generate some images

That process is most of the fun of learning how to paint with your AI tools and what each one can do. It deserves a separate tutorial, so I'm just going to share what I use because I've already done a lot of prompt refinemen. I know what key words and settings works with my weights and I'm using Anything V3 for anime style characters as the most versatile and best tuned. I have a notebook with templates for various image styles. I can drop in an idea in a template and just refine as needed for what extra I want to see. I put the prompt up front, then for anime style images (AnythingV3 or WaifuDiffusion1.3) I add danbooru tags to speficy the composition with extra weight on them (++). The keywords in square braces are all negative prompt elements, things the AI should avoid. The last section includes some keywords that work well to constrain and guide style for me lately. Every image is not a winner. I usually generate 4 at a time and may get one I like after 4, or may need to generate some more.  I ran TXT2IMG to generate candidates at 768x576 using AnythingV3 FP32 weights with prompt "A serene and tranquil florist, with short, blonde hair and piercing, blue eyes, arranges a bouquet of beautiful and fragrant flowers, her movements slow and graceful as she places each blossom with care and attention to detail, her peaceful and nurturing demeanor calming. (1girl solo cleavage looking_at_viewer holding_flowers)++ [poorly_drawn, lowres, extra_legs, extra_arms, bad_hands, bad_feet, bad_anatomy, bad_proportions, anatomical_nonsense, censored, text, signature, logo, watermark, greyscale, monochrome, blurry_subject, redeye] (Incredibly detailed photorealistic body. ultrasharp 8k hires. dynamic composition. cinematic lighting. full color."

4 Refine your favorites

I usually start this by generating variations of the chosen image. If it's not perfect, I may generate some at the same resolution with higher variance to see how it changes. If I need to change one specific thing, I'll use the unified canvas in InvokeAI to mask and regenerate patches with localized prompts. Once I have the composition I like, I generally scale it up. In this case, I liked the composition and I just generated four upscaled candidates to fill in details with a modified text. I dropped the initial prompt and added more DB tags for features I wanted the AI to pay attention to when refining.  I used IMG2IMG to redraw selected image to 2048x1536 using AnythingV3 FP16 weights with prompt "(1girl solo cleavage looking_at_viewer eye_focus, fine_fabric_detail, caustics bokeh, sharpened, enhanced clarity, specular highlights)++ [poorly_drawn, lowres, extra_legs, extra_arms, bad_hands, bad_feet, bad_anatomy, bad_proportions, anatomical_nonsense, censored, text, signature, logo, watermark, greyscale, monochrome, blurry_subject, redeye] (Incredibly detailed photorealistic body. ultrasharp 8k hires. dynamic composition. cinematic lighting. full color.)"  Final upscale with ESRGAN 2x @75% strength (It doesn't actually add detail but with anime style images it sharpens edges and refines lines) and I'm done.

Description
How to create AI Art like me

1 Set up InvokeAI and run it.

If you're comfortable with getting python projects from github and building environments with conda or pip in virtualenv, you can do it yourself. It's not ready for mainstream otherwise. Check out invoke-ai.github.io/InvokeAI/ and follow their instructions to set up.

2 Get some ideas.

Sometimes I bring my own but I like variety and am creating cheap, fast, ephemeral art for my own enjoyment so I generate ideas and pick what I like. I have used various art prompt sites and made my own roll tables. Today I'm enjoying using a ChatGPT session primed to suggest art prompts. I primed the ChatGPT session with "I want you to act as an AI artists muse. You will help artists create artwork using by suggesting creative and interesting image prompt ideas for according to whatever theme I ask for. My first request is "I need prompts that feature mushroom girls having fun.""  Some back and forth later, ChatGPT and I narrowed what I wanted to create, and I selected the prompt "A serene and tranquil florist, with short, blonde hair and piercing, blue eyes, arranges a bouquet of beautiful and fragrant flowers, her movements slow and graceful as she places each blossom with care and attention to detail, her peaceful and nurturing demeanor calming."

3 Generate some images

That process is most of the fun of learning how to paint with your AI tools and what each one can do. It deserves a separate tutorial, so I'm just going to share what I use because I've already done a lot of prompt refinemen. I know what key words and settings works with my weights and I'm using Anything V3 for anime style characters as the most versatile and best tuned. I have a notebook with templates for various image styles. I can drop in an idea in a template and just refine as needed for what extra I want to see. I put the prompt up front, then for anime style images (AnythingV3 or WaifuDiffusion1.3) I add danbooru tags to speficy the composition with extra weight on them (++). The keywords in square braces are all negative prompt elements, things the AI should avoid. The last section includes some keywords that work well to constrain and guide style for me lately. Every image is not a winner. I usually generate 4 at a time and may get one I like after 4, or may need to generate some more.  I ran TXT2IMG to generate candidates at 768x576 using AnythingV3 FP32 weights with prompt "A serene and tranquil florist, with short, blonde hair and piercing, blue eyes, arranges a bouquet of beautiful and fragrant flowers, her movements slow and graceful as she places each blossom with care and attention to detail, her peaceful and nurturing demeanor calming. (1girl solo cleavage looking_at_viewer holding_flowers)++ [poorly_drawn, lowres, extra_legs, extra_arms, bad_hands, bad_feet, bad_anatomy, bad_proportions, anatomical_nonsense, censored, text, signature, logo, watermark, greyscale, monochrome, blurry_subject, redeye] (Incredibly detailed photorealistic body. ultrasharp 8k hires. dynamic composition. cinematic lighting. full color."

4 Refine your favorites

I usually start this by generating variations of the chosen image. If it's not perfect, I may generate some at the same resolution with higher variance to see how it changes. If I need to change one specific thing, I'll use the unified canvas in InvokeAI to mask and regenerate patches with localized prompts. Once I have the composition I like, I generally scale it up. In this case, I liked the composition and I just generated four upscaled candidates to fill in details with a modified text. I dropped the initial prompt and added more DB tags for features I wanted the AI to pay attention to when refining.  I used IMG2IMG to redraw selected image to 2048x1536 using AnythingV3 FP16 weights with prompt "(1girl solo cleavage looking_at_viewer eye_focus, fine_fabric_detail, caustics bokeh, sharpened, enhanced clarity, specular highlights)++ [poorly_drawn, lowres, extra_legs, extra_arms, bad_hands, bad_feet, bad_anatomy, bad_proportions, anatomical_nonsense, censored, text, signature, logo, watermark, greyscale, monochrome, blurry_subject, redeye] (Incredibly detailed photorealistic body. ultrasharp 8k hires. dynamic composition. cinematic lighting. full color.)"  Final upscale with ESRGAN 2x @75% strength (It doesn't actually add detail but with anime style images it sharpens edges and refines lines) and I'm done.