On the LLM side, Min-P sampling gained broad adoption across Inference Engines (vLLM, SGLang, llama.cpp) by scaling the probability cutoff relative to the top token's confidence, allowing higher creativity without descending into incoherent tail tokens.
Sampling & Schedulers
Also known as: Samplers · Temperature · Top-P · Euler / DPM++
Plain English
The math routines that turn raw model predictions into actual outputs—controlling creativity vs. determinism in LLMs (temperature, top-p) and controlling how noise is stepped down into pixels in image and video models (Euler, DPM++, FlowMatch).
Technical Overview
In LLMs, token decoding applies temperature scaling, top-k/top-p (nucleus), or min-p filtering to next-token logits. In Diffusion Models and flow-matching models, ODE/SDE numerical solvers (Euler, Heun, DPM++ 2M Karras, UniPC, FlowMatchEuler) paired with a noise schedule dictate how many steps are needed to converge from Gaussian noise to a sharp latent in ComfyUI.
Chronological Entries (2)
In visual generation, choosing the right sampler + scheduler pair (such as DPM++ 2M with Karras sigmas for SDXL or Euler + Simple/Beta for flow-matching FLUX) can cut generation steps from 50 to 15–20 with identical or better detail.
Direct Relations (3)
- Part ofDiffusion Models
- UsesClassifier-Free Guidance (CFG)
- UsesComfyUI