Seeds
1. The ‘same’ image across different seeds
2. Same seed across different settings
Img-2-Img
Denoising strength x LoRA weight
CFG scale
CFG x textual inversion training steps
Older, with no explanations:
Making flat images 3d with a textual inversion embedding
more coming soon(ish)!
Seeds
1. The ‘same’ image across different seeds
2. Same seed across different settings
Img-2-Img
Denoising strength x LoRA weight
CFG scale
CFG x textual inversion training steps
Older, with no explanations:
Getting around with img2img, textual inversion and controlnet
Making flat images 3d with a textual inversion embedding
Making flat images 3d with a textual inversion embedding
more coming soon(ish)!
This section is still under construction. There’s tons of stuff I’d like to add examples of, but haven’t had time yet. So check back later!
![]()
You can see how much of an effect using a different seed has by looking at the columns (each column is one seed). Changing seeds is it a like rolling the dice to create an entirely new visual universe. One that follows exactly the same rules as all of the others, but just starting with slightly different initial matter.
A different seed with all the same settings will be like you landed up in the equivalent area of a different universe (there are more than 4 billion different seeds to choose from). The rest of the settings determine where in this one instance of the model’s visual universe you land up in.
Changing which sampler is used changes the ‘denoising’ process that generates the image. This means that you change the coordinates of the where you end up in the model, which determines what your output image looks like.
Changing the sampler usually has less impact on the images that I generate than changing seeds. But if you are trying to find a version of an image that is very similar, then experimenting with this can be very helpful.
Four seeds (columns) and four samplers (rows), all other settings remaining the same. Samplers in order: Euler a, DPM2 Karras, DPM++ 2M Karras, DDIM
You can see how much of an effect using a different seed has by looking at the columns (each column is one seed). Changing seeds is it a like rolling the dice to create an entirely new visual universe. One that follows exactly the same rules as all of the others, but just starting with slightly different initial matter.
A different seed with all the same settings will be like you landed up in the equivalent area of a different universe (there are more than 4 billion different seeds to choose from). The rest of the settings determine where in this one instance of the model’s visual universe you land up in.
Changing which sampler is used changes the ‘denoising’ process that generates the image. This means that you change the coordinates of the where you end up in the model, which determines what your output image looks like.
Changing the sampler usually has less impact on the images that I generate than changing seeds. But if you are trying to find a version of an image that is very similar, then experimenting with this can be very helpful.
Still under construction!
There’s tons of stuff I’d like to add examples of, but haven’t had time yet. So check back later!
Four seeds (columns) and four samplers (rows), all other settings remaining the same. Samplers in order: Euler a, DPM2 Karras, DPM++ 2M Karras, DDIM
You can see how much of an effect using a different seed has by looking at the columns (each column is one seed). Changing seeds is it a like rolling the dice to create an entirely new visual universe. One that follows exactly the same rules as all of the others, but just starting with slightly different initial matter.
A different seed with all the same settings will be like you landed up in the equivalent area of a different universe (there are more than 4 billion different seeds to choose from). The rest of the settings determine where in this one instance of the model’s visual universe you land up in.
Changing which sampler is used changes the ‘denoising’ process that generates the image. This means that you change the coordinates of the where you end up in the model, which determines what your output image looks like.
Changing the sampler usually has less impact on the images that I generate than changing seeds. But if you are trying to find a version of an image that is very similar, then experimenting with this can be very helpful.