Termini del modello e dei parametri AI

An AI hallucination is output a model presents confidently even though it is wrong or invented. In text that means a fabricated fact; in images and video it means six-fingered hands, unreadable text on signs, physics that does not hold, and detail an upscaler adds that was never in the source.

hallucination · ai hallucinations · model hallucination · hallucinate

An autoregressive model produces its output one element at a time, each new element conditioned on everything already emitted. Language models work this way, and so does a growing set of image and video models that predict picture tokens or frames in sequence rather than denoising a whole canvas in parallel.

ar model · autoregressive · next token prediction · autoregressive image model

CFG scale is the strength of classifier free guidance, the setting that decides how far a diffusion model is pushed toward your prompt and away from what it would have drawn unprompted. Low values give loose, soft results, high values give literal, contrasty ones, and every model family has its own working range.

what is cfg scale · guidance scale · classifier free guidance · cfg · prompt strength

A CLIP model is a pair of encoders, one for text and one for images, trained so that a caption and its picture land in the same place in a shared embedding space. Image generators use its text encoder to turn a prompt into the numbers that steer generation.

clip · text encoder · contrastive language-image pre-training · clip score · t5 encoder

A consistency model is trained so that any point along a noise-to-image trajectory maps straight to the same endpoint, which lets it finish a generation in one to four steps instead of dozens. It is the distillation technique behind every fast, turbo, flash and lightning tier you see on a model catalog.

model distillation · latent consistency model · lcm · step distillation · turbo model

ControlNet is an add-on network that conditions a diffusion model on a structural map extracted from a reference image, such as a pose skeleton, a depth map, or an edge outline. The prompt still decides the style and content, but the layout is pinned to the map you supply.

what is controlnet · ip adapter · control net · openpose controlnet · depth map conditioning

Denoising strength decides how far your input image is pushed back into noise before a diffusion model starts to denoise it again, which sets how much of the original survives. Sampling steps decide how many passes that rebuild takes. Together they are the two dials behind every image to image edit.

denoising strength · sampling steps · denoise strength · image strength · steps

A diffusion model generates by starting from random noise and removing a little of it at a time, predicting at each step what the image would look like with less noise, guided by your prompt. Training teaches it to reverse a gradual noising process, and that step-by-step structure is what makes the output steerable.

what is a diffusion model · how do diffusion models work · latent diffusion · video diffusion · diffusion vs gan · denoising diffusion

A diffusion transformer is a diffusion model whose denoising backbone is a transformer rather than a convolutional UNet. The latent is cut into patches, each patch becomes a token, and every token attends to every other one at every layer. That single change is what gave current models legible text, long prompts that hold, and video measured in seconds instead of frames.

dit · dit model · mmdit · transformer diffusion backbone

Fine tuning AI models means continuing training a pretrained model on your own images so it learns a subject, style, or product it did not know before. It changes the weights, which is what separates it from prompting and from reference images, and it is the last resort rather than the first.

dreambooth · textual inversion · checkpoint model · stable diffusion checkpoint · custom model training

Flow matching is a training objective that teaches a model a velocity field carrying noise to data along nearly straight paths. Rectified flow is the straight-line version that current image and video families are built on. Because the path is straight, sampling needs far fewer steps than a classic diffusion schedule to reach the same quality.

rectified flow · flow matching model · velocity prediction · rf

A foundation model is a large model pretrained on broad data and then adapted to many downstream jobs rather than built for one. In generative media it is the base checkpoint an entire ecosystem hangs off, since fine-tunes, LoRAs, control adapters, distilled fast variants and task specific versions are all built against a particular base.

base model · foundation models · pretrained model · base checkpoint

A GAN (Generative Adversarial Network) is a pair of neural networks trained against each other, a generator that produces candidate images and a discriminator that judges whether they look real. The contest drives quality up in a single forward pass, which is why GANs remain standard for upscaling and face restoration even though diffusion replaced them for open-ended synthesis.

gans · generative adversarial network · generative adversarial networks · adversarial network · esrgan · stylegan

Generative AI is any model that produces new content (images, video, audio, text) by sampling from a probability distribution it learned during training, rather than retrieving or editing an existing file. Because every run draws a fresh sample, the output changes even when your input does not.

generative ai · gen ai · genai · generative artificial intelligence

Latent space is the compressed coordinate system a generative model works inside, where an image is a few thousand numbers instead of millions of pixels. Generation is a path through that space, decoded back to pixels at the end, which is why a seed is reproducible and why two nearby points look like relatives.

what is latent space · latent · latent representation · latent vector · latents

A LoRA (Low-Rank Adaptation) is a small add-on file that shifts a base model's behaviour toward one specific subject, style or concept without retraining the model itself. It is typically 10 to 300 MB against a multi-gigabyte checkpoint, and it is the standard way to lock a recurring character or house look across many shots.

what is a lora · lora model · lora training · lora fine tuning · what is lora ai · lora vs checkpoint · low-rank adaptation

Multimodal AI describes models that read or write more than one kind of data, usually some mix of text, images, video and audio held in a shared representation. In creative tools it is the reason you can hand a model a reference frame plus a written note and have it understand both at once instead of one after the other.

multimodal model · multimodal models · omni model · vision language model

A negative prompt is a second piece of text a diffusion model is steered away from, the mirror image of the prompt it is steered toward. It reliably suppresses styles, materials, and recurring artefacts, and it is unreliable at removing specific objects from specific places.

negative prompting · negative prompt examples · undesired prompt · exclude prompt

Prompt engineering is the practice of writing and structuring model input so the output lands where you want it. For image and video models it is mostly about word order, specificity and weighting, because these models do not follow instructions the way a chat model does; they match a description.

what is prompt engineering · prompt template · few shot prompting · system prompt · prompt weighting · prompting

A seed number is the integer that initialises the random noise a diffusion model starts from. Hold the seed and every other setting fixed and you get the same output again, which is what turns generation from a slot machine into a controlled experiment where one variable changes at a time.

image seed · random seed · seed value · fixed seed

A tokenizer is the component that splits text into the discrete units a model can process, then maps each unit to an id that becomes a vector (an embedding). It sets the hard limit on how much of your prompt is read at all, and it decides how unusual words get broken apart, which is why rare names often generate something unrelated.

embedding ai · tokenization · tokens · text encoder · clip tokenizer

A transformer model is a neural network that treats its input as a set of tokens and uses an attention mechanism to decide which tokens influence which. It powers large language models, and since the shift to diffusion transformers it powers most current image and video generators too.

attention mechanism · self-attention · transformer architecture · dit · diffusion transformer

A UNet is the convolutional encoder and decoder, joined by skip connections, that predicts the noise to remove at each denoising step of a diffusion model. It was the standard image generation backbone from 2022 to 2024, and its shape is why LoRA and ControlNet exist in the form they do.

u-net · unet architecture · denoising unet · unet diffusion

A VAE (variational autoencoder) is the compressor bolted to both ends of a diffusion model. Its encoder turns pixels into a small latent grid cheap enough to denoise, and its decoder expands the finished latent back into an image or into video frames. Most of the detail flaws you notice at 100% zoom are made here, not by the model that did the generating.

vae stable diffusion · variational autoencoder · vae decoder · latent encoder