Sampler Detailed Explanation
ComfyUI currently has samplers
Old Sampler
DPM Sampler
New Sampler
- heun
- heunpp2
- dpmpp_2
- dpmpp_2_ancestral
- lms
- dpm_fast
- dpm_adaptive
- dpmpp_2s_ancestral
- dpmpp_sde
- dpmpp_sde_gpu
- dpmpp_2m
- dpmpp_2m_sde
- dpmpp_2m_sde_gpu
- dpmpp_3m_sde
- dpmpp_3m_sde_gpu
- ddpm
- lcm
- ddim
- uni_pc
- uni_pc_bh2
- Samplers with names containing ‘a’ or ‘ancestral’
- a is an abbreviation for
ancestral (ancestral sampler)
in English, such samplers will continuously add noise during the sampling process, making each step of the sampling process generate images with a certain degree of randomness, which means non-convergence
- Samplers with ‘GPU’ in their names
- Samplers like
dpmpp_sde_gpu
are optimized for GPU hardware, they can efficiently run large-scale parallel computing tasks on graphics processors, significantly speeding up image generation.
Scheduler
The scheduler controls the denoising process, determining the number of denoising steps and the denoising intensity of each step. The main schedulers in ComfyUI are
- norlmal
- karras
Rendering time is consistent, but after 8 sampling steps, there will be fewer noise points
- exponential
The image will be smoother, the background will be cleaner, but the drawback is that it will lose some details
- sgm_uniform
- simple
- ddim_uniform