A Beginner's Guide to VAEs, CLIP Encoders and Samplers in AI Video Workflows

Radyschen · reddit · 2026-08-29

A beginner-friendly walkthrough of AI video workflow internals, using MiniMax H3 as the example:

VAE: compresses images/video/audio into an abstract format the model understands; decoding recreates detail with some loss, so output isn't 1:1.
CLIP/text encoder: named after OpenAI's original CLIP; today often modified LLMs that convert prompts into the format the video model was trained on — many also understand images and can encode prompt-image relations.
positive output: positive conditioning, an abstract package of all encoded instructions, unreadable by humans.
latent output: the empty 'canvas' sized by frames and resolution.
Sampler & noise: the seed sets the random starting point (same seed, same result); the model was trained to guess what to change in noised real videos, removing a bit of randomness per step.
sigmas: roughly per-step noise percentages, with steps+1 values (the extra one is the start point).

More from Multimodal