The AI: Samplers in the Stable Diffusion.
Lets check 3 samplers, DDIM, Euler and DPM++.
Short Description
DDIM - Denoising Diffusion Implicit Models - were one of the early efficient methods of sampling diffusion models. At the time, most other approaches used an adversarial approach in which one model tries to trick another making them both stronger (GAN neural network architecture).
Euler - Refers to Euler’s method which is a classical approach to solving ODEs.
DPM++ - Diffusion Probabilistic Model-Solver++ is a relatively recent advancement over DDIM which claims to have faster convergence (in terms of number of steps) to create quality images. DPM++ is a higher-order solver which means rather than learning “the way something changes with something else” it is learning “the way that the way that something changes when something else changes”.
Lets make some Pictures
Lets check the differences.
Prompt
dystopian city beautiful landscape, bright luminous night, woman, very intricate, very detailed, sharp, bright, colorful, anime_retro
Sampling Steps 10
CFG 7
Seed 10000
Sampler: dpmpp_2m_sde
Model: technoRealism_v10
Output:
Sampling Steps 30
CFG 7
Seed 10000
Sampler: dpmpp_2m_sde
Model: technoRealism_v10
Output:
Euler
Sampling Steps 30
CFG 7
Seed 10000
Sampler: euler
Model: technoRealism_v10
Output:
DDIM
Sampling Steps 30
CFG 7
Seed 10000
Sampler: DDIM
Model: technoRealism_v10
Output