aDDont
patreon
How do I make CGs
🕑 Added 2024-04-30 16:02:52 +0000 UTCHi, everyone!
Longread ahead :)
A lot of people asked about CGs generation, so here is how I do it. I won't tell how to install stable diffusion and build your setup - you can find dozens of tutorials on how to do it on youtube and civitai.
I make an example out of one of the CGs I need to make for the next update. Mila and Paul are on a date in park. There will be a scene where they are eating. First of all we need to generate a base image.
Here is my prompt and parameters:
pink strapless top, very short shorts, standing, outdoors, evening, golden hours, sunset, blue eyes, makeup, pov date, park, portrait, at table, eating, taco, (score_9, (score_8_up:1.1), 2.5d, source photo, thick outlines, extremely detailed, realistic body proportions, sharp focus, HD, fstop, blurry_background), Expressiveh, <lora:Expressive_H-000001:0.5>, <lora:Smooth Style 2 SDXL_LoRA_Pony Diffusion V6 XL:0.5>, (small breasts|medium breasts, long straight ginger hair, freckles, 30yo petite woman, body freckles, cute face), <lora:mila_sdxl_vanila:0.75>
Negative prompt: shadow, score_4, monochrome, greyscale, worst quality, pony, 3d, frame, borders, blurry face,, milf, plump, fat, big breasts, huge breasts, big ass, huge ass, pregnant, tattoo, pubic hair, short_hair, thick thighs, wavy hair, messy hair, puffy lips, thick lips, loli, kid, child
Steps: 40, Sampler: DPM++ 2M, Schedule type: Karras, CFG scale: 4, Seed: 1171247848, Size: 1152x768, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, Clip skip: 2, Lora hashes: "Expressive_H-000001: 5671f20a9a6b, Smooth Style 2 SDXL_LoRA_Pony Diffusion V6 XL: 93df3fd8fedc, mila_sdxl_vanila: 188ca32feae2", Version: 1.9.3
I make from 10 to 20 images to find the ones that resemble the thing I want to see better. If I see that generation goes in wrong direction I stop it and change something in prompt, or add a control net with reference image (reference, scrible, depth, lineart, ip-adapter or openpose - one of that or combination).
In that case it would be better to delete "eating" tag because it ruins the composition. But I got lucky and get this one:
Let's work with it.
As you can see there are some problems:
Her top isn't pink. Also it should be shorter.
Ratio isn't right (it should be 16:9 but I wouldn't recommend generating base picture far from square format (3:2 is the limit where limbs are decent).
Her hands are weird
Her face is a mess
Benches in the background are weird
Her hair should be a bit longer
So after that I use photoshop.
Ok. We don't need to add people to the background yet, we just need to fill the gaps in correct tone. If denoising str is high enough (more than 0.6) AI should understand what to do with this empty space. Also i deleted taco from her mouth.
Now lets send this to the img2img tab. Here is new generation params:
pink strapless top, very short shorts, outdoors, evening, golden hours, sunset, blue eyes, makeup, pov date, park, portrait, at table, people in background, taco, sitting, (score_9, (score_8_up:1.1), 2.5d, source photo, thick outlines, extremely detailed, realistic body proportions, sharp focus, HD, fstop, blurry_background), (small breasts|medium breasts, very long straight ginger hair, freckles, young petite girl, body freckles, slim, skinny, cute face), <lora:mila_sdxl_vanila:0.75>, Expressiveh, <lora:Expressive_H-000001:0.5>, <lora:Smooth Style 2 SDXL_LoRA_Pony Diffusion V6 XL:0.5>
Negative prompt: shadow, score_4, monochrome, greyscale, worst quality, pony, 3d, frame, borders, blurry face,, milf, plump, fat, big breasts, huge breasts, big ass, huge ass, pregnant, tattoo, pubic hair, short_hair, thick thighs, medium_hair, wavy hair, messy hair, puffy lips, thick lips, futanari,
Steps: 40, Sampler: DPM++ 2M, Schedule type: Karras, CFG scale: 4, Seed: 1586307817, Size: 1920x1080, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, Denoising strength: 0.6, Clip skip: 2, Lora hashes: "mila_sdxl_vanila: 188ca32feae2, Expressive_H-000001: 5671f20a9a6b, Smooth Style 2 SDXL_LoRA_Pony Diffusion V6 XL: 93df3fd8fedc", Version: 1.9.3
And here is the result I like the most:
I generate from 5 to 10 images here. And if the results are good enough I work with the results. If the results are mess I use controlNet, or inpaint in bit by bit.
As you can see prompt changed a bit. I added "people in background" and deleted "eating".
Guess what's next? :) Yeah, Photoshop again. Now we need to fix stuff that will confuse AI in the future generations - additional/missing fingers, morbid limbs etc. On this picture almost everything is good, I'll just fix fingers and ears a bit. Here is the result:
And now we come to the inpainting part. There are following types of inpaintg:
Delete stuff I don't need - use lama cleaner extension. I use to create images with the same background.
"Outpainting" (changing ratio if you are lazy or have bad results with hand drawn outpainting) - use controlNet + "inpaint only + llama"
Whole picture - I use it to change posture or some small parts of the image that need more context.
Only masked - I use it to inpaint faces and sometimes hands and feet. Works wonders if you use it right.
Here we need "only masked" type. Lets start with the right hand. We won't change prompt, just lower the denoising strength to 0.5. If you draw hand correctly it should clean it without messing the gesture.
I also change resolution to 1024 to 1024 and generate two pictures ar once - saves time. If hand is all messed up after that - there are two possible consequences:
You draw it terribly. Redraw the base. Sometimes it's usefull to use "photobashing" - make a picture of your hand in the gesture you need and paste it in photoshop. Use controlNet with depth map or/and lineart. Also sometimes I just generate hands closer, cut it, paste in photosop and use it as a base.
Denoising strength is too high - try lowering it or use "x/y/z plot" to generate it iteratively.
I made the first image but so i redraw it couple times in photoshop untill I got this:
Now let's fix second hand:
Also I don't like the flow of her hair from the right. Let's fix that too. More denosing str and whole picture inpainting (it needs context of light and body):
Now it looks less artificial. Let's fix face next. The hard part is - we should fix her neck near the hand. I hope it won't mess the hand. Only masked inpainting with denoising str 0.6 or so:
I added "loose hair strand" to prompt to guarantee that it stays there. Also when I need to generate couple emotions on the same CG I clear the prompt and use only the emotion tags to guarantee it weights. The result looks good to me.
I fixed it a bit in photoshop (ear looked off, and some small touches here and there).
Let's fix background people and we done. I use only masked inpainting with this prompt:
outdoors, evening, golden hours, sunset, pov date, park, at table, people in background, sitting, (score_9, (score_8_up:1.1), 2.5d, source photo, thick outlines, extremely detailed, realistic body proportions, sharp focus, HD, fstop, blurry_background), Expressiveh, <lora:Expressive_H-000001:0.5>, <lora:Smooth Style 2 SDXL_LoRA_Pony Diffusion V6 XL:0.5>
Negative prompt: shadow, score_4, monochrome, greyscale, worst quality, pony, 3d, frame, borders, blurry face,
Steps: 40, Sampler: DPM++ 2M, Schedule type: Karras, CFG scale: 4, Seed: 2221849866, Size: 1024x1024, Model hash: 67ab2fd8ec, Model: ponyDiffusionV6XL_v6StartWithThisOne, Denoising strength: 0.55, Clip skip: 2, Mask blur: 8, Inpaint area: Only masked, Masked area padding: 32, Lora hashes: "Expressive_H-000001: 5671f20a9a6b, Smooth Style 2 SDXL_LoRA_Pony Diffusion V6 XL: 93df3fd8fedc", Version: 1.9.3
And here is the final result:
