AI video tutorials are full of jargon. One guide tells you to lower the CFG, another to swap the sampler, a third to add a LoRA and check your VRAM. This glossary explains the 40 terms you will meet most often, in plain English, grouped by topic so related ideas sit together. Bookmark it and come back whenever a guide loses you.
Core concepts
- Model: the trained AI system that generates the video. Different models have different strengths, styles and hardware needs.
- Weights: the file, often several gigabytes, that contains what the model learned. “Open weights” means you can download and run it yourself.
- Checkpoint: a saved version of a model’s weights. Fine-tuned checkpoints are versions trained further on a specific style.
- Diffusion: the method most video models use. They start from random noise and remove it step by step until a clip appears.
- Latent space: a compressed representation of video where the model actually works. It is much smaller than full frames, which makes generation possible on normal hardware.
- VAE (variational autoencoder): the component that compresses frames into latent space and decodes them back into pixels.
- Text encoder: the part that turns your prompt into numbers the model can follow.
- Transformer: the type of neural network many recent video models use to relate every part of the clip to every other part, across space and time.
Inputs and modes
- Prompt: your text description of the shot.
- Negative prompt: a list of things you want the model to avoid, such as “blurry, distorted hands, text”.
- Text-to-video (T2V): generating a clip from text alone.
- Image-to-video (I2V): animating a still image.
- Video-to-video (V2V): restyling or transforming an existing clip.
- First and last frame: giving the model a start and end image so it creates the motion in between.
- Reference image: a picture used to guide style or identity without being the literal first frame.
Generation settings
- Seed: the number that sets the starting noise. Same seed plus same settings equals the same result.
- Steps: how many denoising passes the model makes. More steps take longer and can add detail, with diminishing returns.
- CFG / guidance scale: how strongly the model follows your prompt. Too low ignores you; too high looks harsh and over-processed.
- Sampler / scheduler: the mathematical recipe for removing noise. Different samplers trade speed against detail and stability.
- Denoise strength: in image- or video-to-video, how much the model is allowed to change the input. Low keeps more of the source.
- Motion strength: in some tools, how much movement the model adds.
- Frame count: how many frames are generated. With frame rate, it sets the clip length.
- Frame rate (fps): frames per second. Many models generate at 16 to 30 fps.
- Resolution: the width and height in pixels, for example 1280×720. Higher needs more memory.
- Aspect ratio: the shape of the frame: 16:9 landscape, 9:16 vertical, 1:1 square.
- Batch: generating several variations in one run.
Customisation
- Fine-tuning: training a model further on new examples to change its behaviour or style.
- LoRA (low-rank adaptation): a small add-on file that teaches a model a style, character or motion without retraining the whole thing. Easy to share and combine.
- Trigger word: a word in your prompt that activates a LoRA’s learned concept.
- ControlNet / control signals: extra inputs such as depth maps, edges or body poses that guide composition and movement precisely.
- Workflow: in node-based tools like ComfyUI, the chain of steps from inputs to finished video, saved so you can reuse it.
- Node: one building block in a workflow, such as “load model” or “save video”.
Hardware and performance
- GPU: the graphics card that does the heavy maths. Video generation is almost always GPU-bound.
- VRAM: the memory on the graphics card. It limits which models, resolutions and clip lengths you can run.
- Quantisation: storing model weights with fewer bits, such as 8-bit instead of 16-bit, to save memory at a small cost in quality.
- Offloading: moving parts of the model to system RAM when VRAM runs out. It works but is slower.
- Cloud GPU: a graphics card you rent by the hour from a provider instead of buying one.
Post-production
- Upscaling: increasing resolution with an AI model that adds plausible detail.
- Frame interpolation: generating in-between frames to make motion smoother or create slow motion.
- Temporal consistency: how stable details stay from frame to frame. Poor consistency shows as flicker.
- Artifacts: visual errors such as warping, melting hands, extra limbs or garbled text.
- Lip sync: matching mouth movement to an audio track.
Policy and ethics terms
- Content filter / moderation: automated checks that block certain prompts or outputs. “Uncensored” tools relax or remove them.
- Deepfake: realistic synthetic media showing a real person doing or saying something they did not. Harmful without consent and illegal in many places.
- Watermark / provenance: visible marks or hidden metadata, such as content credentials, that show a video was AI-generated.
- Model licence: the legal terms for using a model’s weights, which may restrict commercial use. See Model Licences Explained.
How to use this glossary
You do not need to memorise every term. For your first month, focus on prompt, seed, steps, CFG, resolution and aspect ratio; those six cover almost every beginner decision. Learn the customisation and hardware terms when you move to local tools. For a practical tour of how these ideas fit together, read How AI Video Generators Work and Seeds, CFG and Steps.
Frequently asked questions
What is the single most useful setting to understand?
The seed. Fixing it lets you test one change at a time, which is how you actually learn what works.
Are LoRAs and checkpoints the same?
No. A checkpoint is a full model; a LoRA is a small add-on that modifies one.
Why do tools use different names for the same thing?
There is no standard yet. Guidance, CFG and prompt strength often mean the same setting.
