ComfyUI Glossary

The memory on your GPU that holds the model and sets the ceiling on what you can run.

Third-party node types that add functionality ComfyUI doesn’t ship with by default.

The in-app menu for installing and updating custom nodes and missing models.

Your saved node graph — every node, setting, and wire — as a reloadable JSON file.

The button that actually runs your workflow and generates the output.

Builds an image entirely from a prompt, starting from random noise instead of a photo.

Redraws an existing picture based on a prompt and a denoise strength setting.

Generates a short video clip directly from a prompt, with no starting image.

Animates a single still photo into a short video clip.

The single-purpose block you connect to others to build a workflow.

The main AI model file that controls your image’s style and subject.

The AI technique that turns random noise into an image, step by step.

The network inside a diffusion model that predicts and removes noise at each step.

The compressed, numerical form your image lives in while it’s being generated.

Converts the model’s internal data into a viewable image, and back again.

The technique vs the trained file — the distinction that trips up most beginners.

The file format almost every checkpoint, LoRA, and VAE ships in today.

A compressed model format with selectable quality levels, for running big models on smaller GPUs.

A model stored with fewer bits of precision to shrink its size and VRAM needs.

A fixed 8-bit precision format that roughly halves a model’s VRAM footprint.

A small add-on file that nudges a checkpoint toward a specific style, without retraining it.

A family of LoRA-like methods — LoHa, LoKr, LoCon — that customize a checkpoint using different math.

An older side-network method for nudging style, mostly replaced today by LoRA.

Trains a whole new checkpoint on photos of one subject, for maximum accuracy.

The specific word you must include in your prompt to activate a LoRA or fine-tuned checkpoint.

The images and captions you gather to train a LoRA on a subject or style.

Writing a text description for each training image so the LoRA learns the right associations.

The node that actually turns noise into your finished image.

The algorithm that turns noise into a finished image, step by step.

Paces how the noise-removal effort is spread across your sampling steps.

How many denoising passes the sampler runs to build your finished image.

Controls how strictly the model follows your prompt versus its own creative freedom.

The starting number behind your image’s randomness — lock it to reproduce a result.

Controls how much of a starting image gets erased and replaced during sampling.

A faster, lower-precision replacement for attention that speeds up generation.

The GPU compiler SageAttention and similar optimizations depend on to run.

Increases resolution using an AI model, or a second, higher-resolution sampling pass.

Two reference images defining exactly how a video clip begins and ends.

A dial for how much a clip moves, versus copying real motion from a reference video.

Reads your prompt and converts it into the conditioning the sampler understands.

Text describing what to avoid, steering generation away from it.

A small custom-keyword file trained to represent a style or subject, without changing model weights.

Forces your image’s structure to follow a reference — a pose, depth map, or outline.

Lets a reference image act as part of your prompt, for faces, characters, and styles.

Techniques for keeping the same character’s face recognizable across generations.

Regenerates a masked area of an image while leaving the rest untouched.

Extends an image past its original edges, generating new content that fits the scene.