Diffusion model
Also known as: diffusion, latent diffusion, denoising diffusion
A kind of generative model that creates images or video by starting from random noise and removing it step by step until a picture appears.
Updated
What it means
Diffusion models learn by watching pictures being destroyed. During training, images are buried under more and more noise, and a Neural network learns to reverse each small step. Once trained, it can start from pure noise and work back to an image that never existed.
Most current image generators and many video ones work this way. They are the main engine behind Generative AI imagery, from art tools to Text-to-video systems.
Latent diffusion, the version behind many popular tools, does the denoising in a compressed Latent space rather than on raw pixels, which makes it much lighter to run.
How it works in practice
A Prompt steers the process. At each step, the model nudges the noise toward something that matches the text, and settings control how strictly it follows. More steps usually mean finer detail and a slower result.
For video, the model has to keep each frame consistent with the next. Objects must stay the same shape and move plausibly, which is far harder than producing a single still and the reason generated clips are usually short.
Editing tools use the same technique for AI upscaling, filling in missing areas and removing objects, all by predicting plausible pixels.
What to watch out for
The output is a guess, not a record. A model can hallucinate extra fingers, warped text or impossible physics, and in sport footage that can mean a board, a ball or a limb that bends wrong.
Enhancing real footage with diffusion tools blurs the line between what happened and what was generated. Viewers and judges may question clips that look too perfect, and a Recruiting video or contest entry is the worst place to invite that doubt.
Image quality depends on the Training data, and the rights to that data are still disputed in several countries. A model that saw few snowboarders or surfers during training renders them less convincingly than city streets or faces.
How RawClip handles it
RawClip does not use diffusion to alter or invent footage. Its AI watches the real recordings and picks moments from them, and every frame in the highlight comes from your own files. Your footage is never used to train AI models.