The terms come in groups. Within each group they follow logic rather than the alphabet; that makes them easier to understand. If you are looking for a specific word, use search on the page.
Generation basics
Prompt. A text description of what you want to get. The main way to control generation. The more precise the prompt, the more predictable the result.
Negative prompt. A description of what should not be in the frame: distorted hands, extra objects, text. The model tries to exclude it. Not all models support it.
Text-to-image. Generating an image from text from scratch. You describe it in words, the model invents the picture.
Image-to-image. Generating a new image based on an uploaded one. The model starts from your picture instead of inventing everything itself.
Text-to-video. The same as text-to-image, but the output is video. A moving shot is born from a text description.
Image-to-video. Bringing a picture to life: the image becomes the first frame and the model builds the motion. It gives a more predictable result than generating video from scratch.
Diffusion. The main method of modern generation. The model starts with random noise and gradually removes what is unnecessary, revealing the picture. More in the article on how AI image generation works.
Latent space. A compressed internal representation of an image that the model works with instead of millions of pixels. It is there for speed: computing in it is many times cheaper.
Token (in the context of a prompt). A piece of text that the model splits your prompt into during processing. Roughly a word or part of a word. Not to be confused with tokens as a service’s internal currency.
Hallucination. A situation where the model confidently generates something meaningless or non-existent: extra fingers, unreadable text, impossible geometry.
Generation parameters
Seed. The number that sets the starting noise. One seed plus one prompt equals the same result. A tool for reproducibility: you lock in a good shot or browse variants.
CFG / Guidance scale. How strictly the model sticks to the prompt. A low value means freedom and creativity, a high one means precision, and too high means an overcooked, unnatural picture. The optimum is usually in the middle.
Steps. How many times the model goes through the noise-removal cycle. More steps mean slightly more detail, but slower and more expensive. Beyond a certain number the difference is barely visible, so there is no point pushing steps to the maximum.
Resolution. The size of the picture in pixels. Higher resolution means more detail and a higher generation price. Drafts are best done at low resolution, the final at high.
Aspect ratio. The proportions of the frame: 16:9 for horizontal, 9:16 for vertical and Reels, 1:1 for square. Set the one you need from the start; cropping a finished shot ruins the composition.
Sampler. The algorithm by which noise removal proceeds. Different samplers give slightly different results and speeds. In most services it is hidden under the hood; advanced platforms let you choose.
Batch. The number of pictures per generation. Handy when you need to quickly get several variants of one prompt and pick the best.
Working with images
Inpainting. Regenerating a selected part of the picture. You select the problem area (for example, a hand) and ask for only that to be redrawn. The main tool against artefacts.
Outpainting. Extending the image beyond its borders. It widens the frame by drawing a logical continuation. Useful for changing the framing or format.
Upscale. Increasing resolution while adding detail. Not just stretching but drawing in fine details so the picture stays sharp at a large size. The final stage.
Reference. A sample image you give the model to set a character’s appearance, a style or a composition. More precise than any text description. The basis of working with character consistency.
Mask. A selected area of the image that the model works with. Used in inpainting: the mask says what to redraw and what to leave.
Face restore. Separate processing that improves the face in a picture: removes distortions, adds sharpness. Useful when everything is fine but the face has drifted slightly.
Artefact. Any unwanted distortion in a picture: extra fingers, garbled text, strange geometry, noise. A consequence of the model repeating patterns without understanding the content.
Consistency. Keeping the same character, style or object across different shots. Achieved with references, not descriptions. The main difficulty in series and video.
Advanced
LoRA. A small add-on to a model that teaches it a specific style, character or object. Lightweight and fast, unlike retraining the whole model. This is how, for example, the same hero is generated repeatedly.
ControlNet. A way to set the exact structure of a frame through an additional input: a person’s pose, an object’s outline, a depth map. It gives control over composition that a prompt cannot achieve.
Checkpoint. A saved version of a trained model. Different checkpoints specialise in different things: photorealism, anime, painting. A term from the world of open models.
Embedding. A representation of text or a picture as a set of numbers the model understands. Through embeddings the model links words with meanings. A technical term, but it comes up often.
Denoising step. One iteration of noise removal. The whole generation is made up of many such steps. See also Steps above.
VAE. The component that compresses a picture into latent space and expands it back. Colour reproduction and fine details of the final image depend on it.
Fine-tuning. Additional training of an existing model on your own data so it does a specific task better. Heavier and more expensive than a LoRA, but it gives a deeper result.
Prompt engineering. The skill of writing prompts so as to get the result you need. Not magic but an understanding of which wording the model reads precisely. A breakdown is in the article on how to write prompts for video.
Video terms
Frame. A single still image. Video is a sequence of frames. The model generates them consistently so the motion is smooth.
FPS (frames per second). How many frames are shown per second. More FPS means smoother motion. The standard for most clips is 24 to 30.
Clip / segment. One continuous piece of generated video, usually 5 to 15 seconds. A long video is assembled from several clips.
Lip sync. Synchronising lip movement with speech. Needed for talking avatars. It requires a model with native sound; quality in Russian is still worse than in English.
Native audio. Sound that the model generates together with the video rather than laid on top. It makes the lips match the speech. It usually doubles the generation price.
Camera movement. How the virtual camera behaves: push-in, pull-out, orbit, pan. It is set in the prompt with special terms. A vocabulary is in the article on prompts for video.
Last-frame chaining. A technique for long scenes: the last frame of one clip becomes the first frame of the next. The character and setting carry over automatically, and the join is invisible.
Motion blur. Blurring of fast-moving objects, as in real video. It adds credibility to dynamic scenes. Set in the prompt.
Start and end frame. Images you set as the first and last frame of a clip. The model invents the motion between them. Handy for precise control.
In short
Most of these terms describe either parameters (seed, CFG, steps, resolution), operations on a picture (inpainting, upscaling, references) or how generation works (diffusion, latent space). Understanding them is useful not for the sake of terminology but to control the result more precisely and not overpay for extra attempts.
If you want to dig deeper into a term, we have separate articles: how generation works, how to write prompts and character consistency.
