RIN
horny raccoon
autistic talentless hack
back to blog
1. improve lighting
your lighting is boring.
so. the lighting of your ai images kind of sucks and it feels like there’s nothing you can do about it right? you prompt night, moody lighting, dark, black, and you still get these greyish images that lack anything resembling perfect blacks. they’re at best kind of bluish grey and at worst they totally ignore the prompt and give you a well lit scene aways. you generate a few more and it's the same thing. you slam some (night:1.6) on that shit and it doesn't really help and now it has a streetlight for some reason. ugh. tl;dr: use img2img in the 65-85% denoise range with literally any dark input image. txt2img is ruining your image’s lighting. okay, so, we can fix this, but let’s talk about why first. when you generate an image you start with random noise. check out this boy. on the left is the noise, on the right is the average of the image.
it's noise.
noisy. notice the average is grey. lovely random noise. perfectly uniformly boring. here’s what happens when we run image generation for a few steps, and beside it the average again.
spooky.
shapes. but again, grey. a bit darker but still not dark, and clearly the scene is just well lit with darkish parts now. how can we expect to get dark scenes out of an image when the generator is fighting simultaneously to turn grey into dark and grey into bright? it’s going to land somewhere vaguely in the middle every time. it’s almost impossible to get true dark with the parameters of a normal txt2img generation because its goal is to create contrast. but we can do it in other ways. we can bob ross this shit. we paint the canvas with an underlay so those tones are prominent when we put stuff on top of it. not exactly, but metaphorically. enter: img2img hi. i am the same shit as txt2img. 100% denoise on img2img is basically txt2img. img2img makes the input image noisy based on the parameters and then uses the same techniques to turn that noise into an image. here is a pretty dark image at 80% denoise, ran for a few steps and then interrupted. on its right is the average.
hello darkness.
now our generator will be pulling light out of dark rather than light and dark out of grey. we can be super edgy or just kind of moody based on how high we set the denoise. let's let this image finish a few times.
my god, believable darkness and unbalanced bats.
amazing. okay. what levers do we have so far? first, our input image. it can be whatever as long as it's not dumb grey. next, denoise value. high denoise: brighter image though still weighted lower than normal low denoise: edgelord nightmare fuel or barely visible nonsense
low denoise raccoon.
but we can extend this further. bob ross doesn’t just paint canvas black. he paints it purple. he draws the whole ass sunset before putting the scene on top of it. we can do this too. it’ll bleed through into the color scheme of our image
puurple.
let's apply this idea differently: add some white splotches to the img2img input. splotches becomes visual interest. a light source, a highlight, a character’s eyes, some color on an otherwise monochrome image.
input, denoise(0.91, 0.79, 0.72).
wowee. no more ai grey average. our ‘night’ prompt can actually do what it’s supposed to. we can have evening orange warmth. sickly green darkness. real shit. note: this guide uses euler a, which is important because euler a adds noise before every step. other samplers do different but still interesting things. that’s all. put txt2img away and click the other tab. thanks.
previous guide
back to blog
next guide
×
‹
›