Generative AI: The 2026 Field Guide

This is a reference glossary for all the technical terms used in the document. Use it as a lookup when you encounter a word you have forgotten.

#

3D VAE: A VAE that compresses both spatial and temporal dimensions of video, used to make video diffusion tractable.→ From images to video: the temporal dimension

A

Agentic chat: Driving generation by talking to an interface that creates and edits the media for you, so the model handles the tool operation.→ Production, part two: voice, sound, dimension, and world
Agentic production: A filmmaking loop in which an AI system carries the planning, generation, consistency, and assembly while the human directs by giving notes rather than operating tools.
Aggregator: A tool or platform that wraps many image and video models behind one interface, so you switch models without switching apps.→ Pulling it all together: the modern stack
Alpha (LoRA alpha): A scaling factor that controls how strongly a LoRA's learned change is applied, effectively its volume knob, tuned separately from rank.→ OpenAI (DALL-E, GPT-Image, and GPT-Image-2)
Alt-text: The description attribute web developers add to images for accessibility, which became the enormous free caption corpus that made text-to-image training possible.→ Where the training data actually comes from
Ancestral sampler: A sampler that injects a little fresh noise at each step, adding variation but making results less reproducible for a given seed.→ The single most important piece of infrastructure: ComfyUI
AnimateDiff: An early, widely used community extension that adds motion to image models to produce short video, notable as a custom node rather than a natively supported model.→ Video and audio in the graph
Animatic: A timed, moving version of a storyboard, a rough cut of a sequence made before real generation or filming.→ The hybrid pipeline, AI inside a live-action production
API (Application Programming Interface): The programmatic endpoint you call to run a hosted model from your own code instead of a web interface.→ The deep history, part one: from 2015 to Stable Diffusion
Application platform: A product that wraps underlying models in user-facing creative tools and workflow, capturing the value of usability and brand. Examples include Krea, Leonardo, and OpenArt.→ The generative imagery stack and where you sit
Arena: A public leaderboard that ranks models from large numbers of blind pairwise human preference votes, which tends to reward average appeal over fitness for a specific job.→ When generation learned to reason
Aspect ratio: The width-to-height proportion of an image or video, such as 16:9 or 1:1.→ Midjourney
Attention: The core operation in a Transformer. Each element, a patch or a token, looks at every other element and decides which ones matter for it, letting the model relate distant parts of an image or prompt.→ The deep history, part one: from 2015 to Stable Diffusion
AudioCraft: Meta's open-source audio research bundle, which packages MusicGen, AudioGen, and the EnCodec codec. Its weights are released under a non-commercial license.→ The music and sound landscape
AudioLM: A Google research model that generates speech and audio by predicting learned audio tokens in sequence, a foundational example of the language-model approach to sound.→ How audio generation actually works
AudioSeal: Meta's audio watermarking method, designed to embed a signal detectable down to short segments and to survive light editing.→ The reckoning: licensing, likeness, and consent
AUTOMATIC1111: An older web-based User Interface (UI) for Stable Diffusion, originally the dominant interface in 2022-2023, still widely used by casual users.→ The single most important piece of infrastructure: ComfyUI
Autoregressive: Generating output one piece at a time, each conditioned on what came before, the way language models write text. Some image and video models work this way instead of by diffusion.→ What is a generative visual model, really?
Avatar: A synthetic on-camera presenter driven by a script and a voice, used for explainer, training, and marketing video.→ Lipsync and audio-visual synchronization

B

Benchmark: A standardized test used to compare models on a task, such as prompt adherence or image quality.→ Orientation: the one-paragraph version, and the one-page version
Blackwell: NVIDIA's GPU generation after Hopper, whose B200 and GB200 chips cut the cost per generated image or token against the previous generation.→ The economics of generative imagery
Build vs buy: The recurring decision of whether to build a capability in-house or pay a vendor for it, the choice that most shapes an operator's cost structure.→ A decision framework for your specific situation

C

C2PA (Content Credentials): An open provenance standard, backed by Adobe, OpenAI, and Microsoft, that attaches cryptographically signed metadata proving a file's origin, though the metadata strips easily when a file is re-encoded.→ Questions readers always ask
Camera control: Prompting or conditioning that directs virtual camera moves such as pan, orbit, dolly, and zoom in a generated video.→ Kuaishou (Kling)
Canny: An edge-detection map, from the Canny algorithm, used with ControlNet to make a generated image follow the outlines of a reference.→ The professional generation workflow
Captioning: Writing the text descriptions that accompany training images. In LoRA training, whatever you caption becomes a controllable variable and whatever you omit becomes baked-in identity.→ The deep history of training datasets
CFG (Classifier-Free Guidance): A technique that pushes a diffusion model to follow the prompt more strongly by contrasting its prediction with and without the prompt. Higher values obey the text more but can look oversaturated or 'fried'.→ The single most important piece of infrastructure: ComfyUI
Character consistency: The problem of generating the same character, with the same face and identity, across many images, poses, and scenes. It is the main bottleneck holding back AI narrative work.→ Black Forest Labs (FLUX)
Character LoRA: A small model trained on images of one specific subject so that a fixed identity can be summoned consistently across an entire production.→ The training dataset is where LoRAs are made or broken
Character sheet: A set of reference images of the same character from several angles and expressions, used to lock identity across generations.→ The workflows opening up right now
Checkpoint: A saved set of model weights. The .safetensors or .ckpt files you download to run a model.→ The single most important piece of infrastructure: ComfyUI
Civitai: A community website for sharing Stable Diffusion and FLUX models, LoRAs, and embeddings.→ The deep history of training datasets
CLIP: A text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward.→ Refinements on the basic recipe
Cold start: The initial delay when a serverless model has to load onto a GPU before it can run, after which subsequent requests are fast.
ComfyUI: The dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org.→ The deep history of training datasets
ComfyUI Manager: The extension that installs, updates, and toggles custom nodes, and whose install-missing feature reads a shared workflow and fetches the nodes it needs.→ The ecosystem: the Manager, custom nodes, and the registry
Common Crawl: A nonprofit that has continuously crawled the public web since 2008 and publishes the archive for free, the raw source most image training datasets are extracted from.→ Where the training data actually comes from
Compositing: Combining multiple image or video layers into one final frame, a core post-production step.→ The single most important piece of infrastructure: ComfyUI
Concept LoRA: A LoRA that teaches or repairs a specific concept, action, or object the base model handles badly, such as correct hands or a particular pose.→ The four main types of LoRAs and what each one is for
Conditioning: Any extra input that steers generation beyond the prompt, such as a depth map, pose, edge map, or reference image.→ The deep history, part two: the transformer takes over
Conditioning map: An image such as an edge, depth, or pose map that constrains generation to a structure. See ControlNet.→ The recipes, part two: LoRA, ControlNet, IPAdapter
Consistency model: A model trained to jump most of the way to a finished image in one or a few steps rather than many. A key idea behind 'turbo' and 'lightning' fast variants.→ Distillation and the race to one-step generation
ControlNet: A conditioning technique that lets you provide structural guidance (depth, edges, pose) to a diffusion model to control spatial layout.→ The single most important piece of infrastructure: ComfyUI
Conversational editing: Refining an image through a sequence of instructions where each builds on the last result, treating the model as a collaborator that holds the image steady between changes. Also called multi-turn editing.→ Google DeepMind (Imagen, Nano Banana)
Convolutional neural network (CNN): A network built from convolutions, operations that look at small local patches of an image, which dominated computer vision before transformers and forms the basis of the U-Net.
Cross-attention: The step where the image being generated attends to the text embedding. It is the mechanism by which the prompt actually steers the picture.→ How diffusion really works, in plain English
Custom node: A community-written extension that adds new operations to ComfyUI. Because it is arbitrary Python running on your machine, custom nodes are powerful and a genuine security consideration.→ The recipes, part two: LoRA, ControlNet, IPAdapter

D

Day-zero model support: An inference platform hosting a newly released model within hours of its launch, so operators can evaluate it immediately.→ The generative imagery stack and where you sit
DDIM (Denoising Diffusion Implicit Model): A sampling method that generates an image in far fewer steps than DDPM by following a deterministic, non-random path through the denoising process.→ The single most important piece of infrastructure: ComfyUI
DDPM (Denoising Diffusion Probabilistic Model): The 2020 method that made diffusion practical for image generation. It defines a forward process that adds noise and trains a network to reverse it step by step. Most modern image models descend from this idea.→ The deep history, part one: from 2015 to Stable Diffusion
De-aging: Visual effects work that makes a performer look younger on screen, now often done with a neural model trained per actor per age.→ Post-production: finishing, and the rights problem
Deepfake: A synthetic image, audio, or video of a real person or event made to look genuine, a category the EU AI Act now requires be disclosed.→ Questions readers always ask
Denoise strength: A parameter from zero to one that sets how much of an image to repaint. At one you generate from scratch, and dropping it below one is the whole trick behind image-to-image and refinement passes.
Denoiser: The neural network at the heart of a diffusion model. Its only job is to look at a noisy image and predict the noise to remove, which is how the image is gradually cleaned into a picture.→ Refinements on the basic recipe
Denoising: The core loop of diffusion. Starting from pure noise, the model repeatedly predicts and subtracts a little noise until a coherent image emerges.→ Orientation: the one-paragraph version, and the one-page version
Depth map: A grayscale image encoding how far each pixel is from the camera, used as conditioning to control 3D structure.→ The single most important piece of infrastructure: ComfyUI
Diffusion model: A generative model that learns to produce images by training a neural network to reverse a noise-adding process. Generates by starting from random noise and iteratively denoising to produce a clean image.→ What is a generative visual model, really?
Digital double: A photorealistic computer-generated stand-in for a real actor, used for stunts, de-aging, or shots a performer cannot physically do.→ The hybrid pipeline, AI inside a live-action production
Digital replica: The industry and union term for an AI-generated reproduction of a real performer's voice or likeness, the thing SAG-AFTRA requires consent to create and use.→ The hybrid pipeline, AI inside a live-action production
Distillation: A technique where a smaller, faster 'student' model is trained to mimic the outputs of a larger 'teacher' model in fewer sampling steps. Used to produce fast variants of slow models (FLUX schnell, SDXL Turbo, etc.).→ How diffusion really works, in plain English
Distribution Matching Distillation (DMD): A distillation method that closed most of the quality gap for few-step image generation, letting a student match a multi-step teacher while running far faster.→ Distillation and the race to one-step generation
DiT (Diffusion Transformer): A diffusion model architecture that uses a Transformer instead of a U-Net. Introduced by Peebles and Xie in 2022. Scales much better than U-Nets and is the basis for almost all current frontier models.→ The deep history, part two: the transformer takes over
DPM++: A family of fast, high-quality diffusion samplers that reach good images in relatively few steps.
DreamBooth: A fine-tuning method that teaches a model a specific subject, such as a person or product, from a handful of images so it can be generated in new scenes.

E

Elo: A rating derived from head-to-head human preference votes, borrowed from chess, used to rank models on leaderboards.→ Google DeepMind (Imagen, Nano Banana)
ELVIS Act: A 2024 Tennessee law, the first to explicitly protect a person's voice, including simulations of it, and to reach the makers of cloning tools.→ The reckoning: licensing, likeness, and consent
Embedding: A list of numbers that represents the meaning of a word, image, or concept in a form the model can compute with. Similar things sit close together in this numeric space.→ Refinements on the basic recipe
EnCodec: Meta's neural audio codec, a refinement of residual vector quantization that became the tokenizer under a wave of speech and music models.→ How audio generation actually works
Enforce-then-license: The studio strategy of suing infringing model makers first and then cutting controlled licensing deals from a position of legal strength.→ The deals, and what actually happened
Enrollment: The longer reference recording, often many minutes, used to build a high-fidelity voice clone that holds up under close listening.→ The audio model landscape, in practice
ESRGAN / Real-ESRGAN: A family of neural upscaler models that enlarge finished pixels and sharpen existing detail, widely used as the classical, faithful end of upscaling.
EU AI Act: The European Union's regulation of AI systems, whose Article 50 transparency rules require disclosing deepfakes and marking synthetic content as artificial, applying from August 2, 2026.→ Questions readers always ask
Euler sampler: A simple, reliable diffusion sampler named after the numerical method it uses. 'Euler a' is its ancestral variant.

F

F5-TTS: A widely used open text-to-speech model that combines flow matching with a Diffusion Transformer, part of the faster non-token lineage of voice synthesis.→ How audio generation fits, in brief
Face restoration: A finishing pass that rebuilds facial detail softened or damaged earlier in a pipeline, for example after an aggressive upscale.→ The single most important piece of infrastructure: ComfyUI
Face swap: Replacing a face in an image or video with another identity, often used as a final correction pass to enforce character consistency.→ The hybrid pipeline, AI inside a live-action production
Fair use: The US legal doctrine that permits some unlicensed use of copyrighted work, and the central defense labs raise for training models on scraped data. Its application to AI training is still unresolved.→ Where the training data actually comes from
FID (Frechet Inception Distance): A metric that scores how close generated images are to real ones in aggregate. Lower is better. It is imperfect but widely reported.→ Distillation and the race to one-step generation
Fine-tuning: Continuing to train an existing model on new data to specialize it for a style, subject, or domain, without training from scratch.→ Google DeepMind (Imagen, Nano Banana)
First-and-last-frame: A video workflow where you fix the opening and closing frame and the model generates only the motion between them.→ The workflows opening up right now
Flat-lay: A product photo taken from directly above with the item laid flat, a common input that AI tools convert into on-model imagery.→ Excellence in fashion ecom and film/TV, today and in the months ahead
Foley: The everyday sound effects of a scene, footsteps, cloth, impacts, now increasingly generated from the video itself.→ When video learned to talk
Forward process: In diffusion, the fixed procedure that gradually destroys an image by adding a little noise at each step until nothing but static remains. It involves no learning.→ The deep history, part one: from 2015 to Stable Diffusion
Frame interpolation: Generating in-between frames to raise a video's frame rate or smooth its motion.→ The single most important piece of infrastructure: ComfyUI
Full fine-tuning: Retraining every weight of a model on new data, which is powerful but needs enormous GPU memory and produces a full-size copy of the model, the expensive alternative that LoRA was invented to avoid.→ What a LoRA actually is
Full-duplex: A voice system that listens and speaks at the same time, handling interruptions and back-channel sounds the way people actually talk, rather than taking strict turns.→ Production, part two: voice, sound, dimension, and world

G

GAN (Generative Adversarial Network): An older family of generative models where two networks, a generator and a discriminator, train against each other. The generator tries to produce realistic images while the discriminator tries to tell real from fake. GANs made sharp images but were unstable to train and largely gave way to diffusion.→ What is a generative visual model, really?
Gaussian splatting: A 3D representation that renders a scene as a cloud of many small fuzzy blobs, one of the export formats world-model tools use to drop generated environments into real production pipelines.
Generation-to-keep ratio: The number of clips or images generated for every one actually used, a measure of the curation labor behind polished AI work.→ The films and the filmmakers
GGUF: A file format for storing quantized models efficiently, common for running large models on limited hardware.→ Alibaba (Qwen-Image, Tongyi Wanxiang)
Ghost mannequin: A product photo of a garment shaped as if worn but with the mannequin or model edited out, a common starting point for on-model generation.
Glaze: A tool from the University of Chicago that cloaks an artist's images so models cannot easily learn and copy their style.→ The recipes, part one: img2img, inpainting, outpainting, upscaling
GPU (Graphics Processing Unit): The parallel processor that runs image and video models. Its speed and memory largely determine what you can generate and how fast.→ Modern engineering: distillation, MoE, and the speed-quality frontier
GPU cloud: A provider that buys GPUs and rents them by the hour or second, sitting between the hardware maker and the inference platforms. Examples include CoreWeave, Lambda, and RunPod.→ The generative imagery stack and where you sit
GPU-hour: One hour of a single GPU's time, the standard unit for quoting the compute cost of training or running a model.→ The economics of generative imagery
Gradient descent: The training procedure that repeatedly nudges a model's parameters in the direction that reduces its error, slowly turning a random network into a useful one.→ What is a generative visual model, really?
Guidance scale (Classifier-Free Guidance (CFG) scale): A parameter that controls how strongly the model adheres to the prompt versus producing diverse outputs. Higher values mean stricter prompt following. Typical range 5-10.→ How diffusion really works, in plain English

H

Hallucination: When a model confidently generates something incorrect or invented, such as garbled text, extra fingers, or a product detail that was never in the prompt.→ Questions readers always ask
Hi-res fix: A two-stage technique that generates a small image, upscales the latent, then runs a second sampling pass at low denoise to invent fine detail at the larger size.→ The recipes, part one: img2img, inpainting, outpainting, upscaling
Hugging Face: The dominant repository for open-source AI models, including all the open-weight image and video models discussed in this document.→ The deep history of training datasets
Hyperparameter: A setting chosen before training that shapes how a model learns, such as a LoRA's rank and alpha or the number of training steps, as opposed to the weights learned during training.
Hyperscaler: One of the largest general-purpose cloud providers, meaning AWS, Google Cloud, and Azure, which rent GPUs as part of a broader offering.→ How research actually works

I

Identity drift: The tendency of a generated character's face, wardrobe, or props to change from shot to shot, the central failure that consistency tools exist to fight.→ Image and video editing models
Image-first workflow: The practice of generating and art-directing a still frame until it is exactly right, then animating that locked image with image-to-video.
Image-to-video: Starting from a still image and generating a video that animates it according to a prompt.→ Midjourney
ImageNet: A large labeled image dataset long used as a standard benchmark for training and comparing computer vision and image generation models.→ The deep history, part two: the transformer takes over
Img2img: Image-to-image generation: starting from an existing image rather than from pure noise, and applying a prompt to modify it.→ The professional generation workflow
Indemnification: A vendor's promise to cover a customer's legal costs if the product triggers a lawsuit, offered by some licensed-data model providers as protection against copyright claims on generated output.→ The deep history of training datasets
Inference: Running a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training.→ What is a generative visual model, really?
Inference platform: A company that turns raw rented GPUs into pay-per-generation APIs, hosting model weights and handling scaling so developers do not manage the hardware. Examples include fal, Replicate, and Modal.→ Inside fal: one platform up close
Infinite canvas: A spatial workspace of nodes and connections where the output of one model feeds the next, making a whole pipeline visible and reusable.→ Production, part one: the engines and the canvas
Inpainting: Editing a specific masked region of an image while preserving the rest. The classic AI editing approach.→ Questions readers always ask
InstantID: A technique for preserving a specific person's identity across generations from a single reference photo, without training a LoRA.→ The workflows opening up right now
Instruction-based editing: Editing an image by giving a text instruction, such as change the background to a beach, and letting the model decide what to change and what to leave alone, without a hand-drawn mask.→ Alibaba (Qwen-Image, Tongyi Wanxiang)
IP-Adapter (Image Prompt Adapter): A method for conditioning generation on a reference image, so a subject or style from a picture carries into new outputs.→ The workflows opening up right now

K

Karras schedule: A popular noise schedule that spaces the denoising steps to get better quality in fewer steps. Often paired with DPM++ samplers.→ The default workflow, node by node
Keyframe: A specific frame you fix as an anchor, for example the first and last frame of a shot, that the model generates toward or between.→ From images to video: the temporal dimension
KSampler: The ComfyUI node where denoising actually happens, taking the model, prompts, and a latent and running the sampling loop under its seed, steps, CFG, sampler, and scheduler settings.→ The single most important piece of infrastructure: ComfyUI

L

LAION: A German nonprofit that assembled the LAION-5B dataset of 5 billion image-text pairs used to train Stable Diffusion.→ The deep history, part one: from 2015 to Stable Diffusion
Latency: The delay between asking for a generation and getting the first result. Distinct from throughput.→ When generation learned to reason
Latent consistency: The application of consistency-model training in latent space to make diffusion models generate in very few steps.→ Krea (Krea-1)
Latent diffusion: Diffusion that operates in a compressed 'latent' space produced by a VAE, rather than directly in pixel space. Dramatically reduces compute cost. Used by essentially every modern image and video model.→ The deep history, part one: from 2015 to Stable Diffusion
Latent space: The compressed numeric representation an image is turned into before diffusion. Working here instead of on raw pixels is what makes modern image models fast enough to run.→ The deep history, part one: from 2015 to Stable Diffusion
Licensed data: Training data used with the permission of and usually payment to its rights holders, the basis for models marketed as commercially clean.→ Where the training data actually comes from
Likeness vault: A managed store of a performer's cleared voice and likeness, built by talent agencies so that consented digital replicas can be licensed.
Lipsync: Synchronizing the mouth movements of a person in video with an audio track of speech.→ Pulling it all together: the modern stack
LM Arena: A public leaderboard where anonymous models compete in blind head-to-head human votes, increasingly the venue where labs quietly test new models before official release.→ OpenAI (DALL-E, GPT-Image, and GPT-Image-2)
LoRA (Low-Rank Adaptation): A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models.→ The deep history of training datasets

M

Matting: Separating a subject from its background to produce a clean cutout or mask, the first step in a lot of compositing.→ Inside fal: one platform up close
MCP (Model Context Protocol): An open standard that lets an AI assistant drive external tools. Applied to ComfyUI, it lets an agent search nodes, assemble a workflow, and run generations on your behalf.→ The single most important piece of infrastructure: ComfyUI
Mean Opinion Score (MOS): The standard audio quality metric, obtained by playing a clip and asking listeners to rate it from one to five. Natural speech scores around 4.5.→ How audio generation actually works
Memorization: When a model can reproduce specific training examples on demand rather than only learning general patterns, a central concern in the copyright debate.→ How generative visual models actually work, told as a story
Mesh: The 3D geometry of an object, a network of vertices and faces, the core output of a 3D generation model.→ 3D generation models
Microdrama: A vertical, phone-first serialized drama told in very short episodes, the format where AI-generated narrative reaches paying audiences first.→ Distribution: where AI-native content lives and pays
Mimi: Kyutai's streaming neural audio codec, designed to generate sound incrementally with low enough latency for live conversation.→ How audio generation actually works
MM-DiT (Multimodal Diffusion Transformer): A variant of DiT that processes text and image (or audio) tokens in separate parallel streams joined by attention. Used by Stable Diffusion 3, FLUX, and most newer models. Improves prompt following significantly.→ Refinements on the basic recipe
Moat: A durable competitive advantage that protects a business from being copied or undercut, such as workflow depth, proprietary data, or brand and community.→ The deep history of training datasets
Mode collapse: A failure, classically in GANs, where the generator produces only a few kinds of output regardless of input, losing diversity. Distinct from model collapse.→ What is a generative visual model, really?
Model collapse: The degradation that happens when models are trained on their own or other models' synthetic output over successive generations, gradually losing quality and diversity. Distinct from mode collapse.
Model lab: A company that trains and ships the frontier generation models themselves, such as Black Forest Labs, Runway, or Google.→ The deep history, part one: from 2015 to Stable Diffusion
Model routing: Automatically sending each request to the best model for it, by quality, speed, or cost, instead of committing to one model for everything.→ Google DeepMind (Imagen, Nano Banana)
Model-first: A platform strategy built around training and improving one's own flagship generation model, exemplified by Runway and Google's Flow.→ Google DeepMind (Imagen, Nano Banana)
MoE (Mixture of Experts): An architecture where multiple specialized subnetworks ('experts') are routed to handle different inputs, allowing larger total parameter counts without proportional compute increases. Used by Wan 2.2 and several language models.→ Modern engineering: distillation, MoE, and the speed-quality frontier
Multi-teacher captioning: A pipeline for building video datasets that runs several vision-language models on each clip to propose captions, then scores and keeps the best one, producing high-quality labels at scale.→ The deep history of training datasets
Multimodal: Handling more than one kind of data, for example text and images together, in a single model.→ Refinements on the basic recipe
MusicGen: Meta's open music generation model, which compresses audio into tokens and predicts them in order, producing instruments and vocals together.→ How audio generation actually works

N

Native audio: Sound generated jointly with video in a single pass, so footsteps, dialogue, and ambience line up with the picture, rather than being added afterward.→ The economics of generative imagery
Negative prompt: A description of things you do not want in the output. The model is steered away from these concepts during sampling.→ Prompting at a senior level
Neural audio codec: A learned compressor that turns a sound waveform into a short sequence of discrete tokens and back, the audio equivalent of the VAE. Its quality sets a ceiling on the quality of everything built on top of it.→ How audio generation actually works
Neural network: A mathematical function built from layers of simple operations, with millions or billions of adjustable numbers inside it, that learns a task by having those numbers tuned on examples.→ Orientation: the one-paragraph version, and the one-page version
Nightshade: A tool from the University of Chicago that subtly poisons images so that models training on them without permission are degraded.→ Slop, consent, and the fight over the human
NO FAKES Act: A proposed United States federal law that would create a nationwide right against unauthorized digital replicas of a person's voice and likeness.→ The reckoning: licensing, likeness, and consent
Node: A single block in a visual pipeline like ComfyUI that does one operation and wires to others to form a workflow.→ The single most important piece of infrastructure: ComfyUI
Node graph: A visual pipeline where each box performs one operation and wires to others, the interface ComfyUI made standard.→ The single most important piece of infrastructure: ComfyUI
Noise schedule: The recipe that decides how much noise is added at each step of the forward process and removed on the way back. It strongly affects quality and speed.→ The single most important piece of infrastructure: ComfyUI
NVIDIA A100 / H100: The data-center GPUs that train and run most large generative models, rented by the hour from cloud providers and often the dominant line item in a model's cost.

O

Omni-modal: Describing a model that can take any mix of text, image, video, and audio as input and produce any mix as output, with the modalities interchangeable rather than fixed.→ MiniMax (Hailuo)
On-model: Product imagery that shows a garment worn by a real or synthetic model rather than laid flat or on a mannequin.→ Excellence in fashion ecom and film/TV, today and in the months ahead
Open weights: Models whose trained parameters are publicly downloadable, allowing anyone to run them locally or fine-tune them. Contrasts with closed/API-only models.→ The economics of generative imagery
OpenPose: A skeleton map of a figure's pose, used as conditioning to control the body position of generated people.→ The professional generation workflow
Opt-in versus opt-out: Whether a person must actively agree before their likeness or voice is used (opt-in) or must actively object to stop it (opt-out), a distinction unions fought over after Sora 2.
Orchestration-first: A platform strategy built around routing across many third-party models behind one workflow, exemplified by OpenArt and the aggregators. Also called workflow-first.→ The platforms and the agentic production loop
Outpainting: Extending an image beyond its original borders by generating new content that fits with the existing edges.→ Image and video editing models
Overfitting: When a model memorizes its training data instead of learning general patterns, so it does well on what it has seen and poorly on anything new. A common risk when training LoRAs.→ The two knobs that matter: rank and alpha

P

Parameter: One individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost.→ The deep history, part one: from 2015 to Stable Diffusion
PBR (Physically Based Rendering): A material standard describing how a surface responds to light (color, roughness, metalness), so a 3D asset looks right under any lighting.→ Inside fal: one platform up close
Performance transfer: Driving a generated character with a real performer's captured head, face, and body motion from a video, rather than from an audio track, as in Runway's Act-Two.→ Lipsync and audio-visual synchronization
Plate: A piece of photographed footage that generated or computer-generated elements have to be matched and composited against.→ The single most important piece of infrastructure: ComfyUI
Previs (Previsualization): Planning a shot or sequence visually before production, so the expensive decisions get made cheaply.→ Development: everything before a frame exists
Programmatic video: Video produced directly from code and data with no human editing a timeline, used for personalized and bulk output at scale.
Prompt: The text description you provide to a model to specify what you want it to generate.→ Orientation: the one-paragraph version, and the one-page version
Prompt adherence: How faithfully a model follows the specifics of your prompt, meaning the objects, counts, colors, spatial relationships, and text you asked for. High adherence means you get what you described, not just something in the right neighborhood.→ How diffusion really works, in plain English
Prompt engineering: The craft of writing and structuring prompts to reliably get the result you want.→ Questions readers always ask
Provenance: A verifiable record of how a piece of media was made and by whom, increasingly required so that synthetic content can be traced.→ Questions readers always ask

Q

Quantization: Shrinking a model by storing its weights at lower numeric precision, for example 8 or 4 bits instead of 16, so it fits in less memory and runs faster, usually with a small quality cost.→ Questions readers always ask

R

Rank (LoRA rank): The inner dimension of a LoRA's two small matrices, which sets how much capacity it has to learn. Higher rank captures more complex behavior but risks overfitting and produces larger files.→ The deep history of training datasets
Realtime generation: Generation fast enough to respond as you type, paint, or move, turning the model from a batch render job into a live instrument.→ Realtime and interactive, a different mode
Rectified flow / flow matching: A training technique that teaches diffusion models to follow straighter trajectories from noise to data, allowing fewer sampling steps and better quality. Used by Stable Diffusion 3, FLUX, and most newer models.
Reference image: An image fed to a model as conditioning to carry a subject, style, or composition into new output, rather than describing it in words.→ Midjourney
Refiner: A second model or pass that takes a near-finished image and sharpens its details, used in some two-stage pipelines.→ Stability AI (Stable Diffusion family)
Relighting: Changing the lighting on a subject after it was captured or generated, one of the harder problems in compositing.→ Runway (Gen-3, Gen-4, Aleph, Act-Two)
Residual vector quantization (RVQ): The technique behind modern audio codecs, which quantizes a slice of sound, measures the leftover error, quantizes that residual again with another codebook, and repeats across several codebooks, giving high fidelity at a low bitrate.→ How audio generation actually works
Resolution: The pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time.→ The deep history, part one: from 2015 to Stable Diffusion
Retopology: Rebuilding a generated 3D mesh with clean, efficient geometry so it can be animated and used in a real production pipeline.→ 3D generation models
Reverse process: In diffusion, the learned procedure that starts from pure noise and repeatedly removes a little predicted noise until a clean image emerges. Generation is the reverse process run to completion.→ The deep history, part one: from 2015 to Stable Diffusion
RLHF (Reinforcement Learning from Human Feedback): Training that tunes a model toward outputs humans prefer, using human ratings as the reward signal.→ How research actually works
Rotoscoping: Isolating a subject from its background frame by frame, traditionally by hand and increasingly automated by AI.→ Excellence in fashion ecom and film/TV, today and in the months ahead

S

Safetensors: A safe, fast file format for storing model weights that, unlike older pickle formats, cannot execute hidden code when loaded.
SAG-AFTRA: The United States labor union representing actors and other media performers, whose agreements require consent and compensation for any digital replica.→ ElevenLabs
Sampler: The algorithm used to perform the denoising steps. Different samplers (Euler, DPM++, etc.) have different speed-quality trade-offs.→ The single most important piece of infrastructure: ComfyUI
Sampling step: One iteration of the denoising loop. Modern models typically use 20-50 steps for full quality, or 4-8 for distilled fast variants.
Scheduler: The algorithm that decides the sequence of denoising steps and how each is applied. Often called the sampler. Different schedulers trade speed against quality.→ The single most important piece of infrastructure: ComfyUI
Section 230: The United States law that shields online platforms from liability for content their users post, whose application to AI-generated output is unsettled.→ Legal and operational realities
Seed: A random number that determines the starting noise for generation. The same seed plus the same prompt produces the same image. Different seeds with the same prompt produce different variations.→ How diffusion really works, in plain English
Seed-locking: Reusing the same random seed across related generations as a cheap way to hold appearance and composition steady.→ The craft: technique, volume, and the edit
Self-attention: Attention within a single sequence, where image patches or words relate to each other, giving the model a sense of global structure.
Semantic negligence: The structural failure of diffusion models to honor negation, because their training captions describe what is present in an image and never what is absent, which is why 'no glasses' often produces glasses.→ The reliability gap: why professional use still hurts
Serverless: Running a model on infrastructure that spins up on demand and bills per use, so you do not manage an always-on GPU.→ Local or rented: where ComfyUI runs
Serverless GPU: A GPU you rent per second or per request rather than renting or owning full time, the model behind most hosted generation.
Set extension: Digitally enlarging a physical set beyond what was actually built, so a small stage reads as a full location.→ The hybrid pipeline, AI inside a live-action production
Shot extension: Generating a new clip that connects to the end of a previous one, a way to build continuous motion past a model's short length ceiling.→ How an AI film actually gets made
Slop: AI content shared without human review or intention, defined less by the tool that made it than by the missing step of someone curating it before it goes out.→ How an AI film actually gets made
SoundStream: Google's 2021 neural audio codec, the model that introduced residual vector quantization for audio.→ How audio generation actually works
Spatiotemporal attention: Attention that operates across both spatial dimensions (within frames) and temporal dimensions (across frames). Essential for high-quality video generation.→ From images to video: the temporal dimension
Speech-to-speech: Converting a recorded performance into another voice while keeping the original timing and emotion, the approach serious film work prefers over pure text-to-speech.→ ElevenLabs
Stem separation: Splitting a finished music track back into its component parts, such as vocals, drums, and bass, which music tools increasingly expose as stems.→ How audio generation actually works
Stems: The separated individual tracks of a piece of music, such as vocals, drums, and bass, which can be edited or remixed on their own.→ Suno
Storyboard: A sequence of frames that plans a film's shots and staging before anything is generated or filmed.→ Headless and agentic: driving the graph by machine
Style LoRA: A LoRA trained on images that share an aesthetic but vary in subject, so it can render any prompt in that art style, photographic look, or brand identity.→ Krea (Krea-1)
Subgraph: A cluster of ComfyUI nodes collapsed into a single reusable super-node that can be nested and shared, the key to keeping large workflows maintainable.→ Becoming senior
Subject LoRA: A LoRA trained to reproduce one specific person, character, or object across new scenes and poses, the most common and easiest kind to train.→ The four main types of LoRAs and what each one is for
Super-resolution: The technical name for increasing image or video resolution while adding plausible detail, the task performed by upscalers.→ Upscaling models
Synthetic data: Training data generated by existing AI models, such as model-made images or machine-written captions, increasingly used to supplement or replace scraped web data.→ Where the training data actually comes from
SynthID: Google's watermarking system, which embeds an identifying signal in images, audio, video, and text it generates.→ Google DeepMind (Imagen, Nano Banana)

T

Talking head: Video of a single person speaking to camera, the format current video models handle most reliably, which is why many AI films are built around it.→ Inside fal: one platform up close
Temporal coherence: The property of video where elements remain consistent across frames. The central challenge of video generation.→ From images to video: the temporal dimension
Text encoder: A neural network that converts text prompts into numerical embeddings that the diffusion model can use. Modern models use combinations including CLIP and T5.→ Orientation: the one-paragraph version, and the one-page version
Text-to-3D: Generating a three-dimensional mesh and textures from a text prompt, for use in games, AR, and virtual production.→ Production, part two: voice, sound, dimension, and world
Text-to-speech (TTS): Generating spoken audio from written text. Modern neural TTS learns the mapping from text to sound directly from data rather than gluing together recorded snippets.→ How audio generation actually works
Text-to-video: Generating video from text alone, with no starting image.→ The deep history, part two: the transformer takes over
Textual inversion: A lightweight training method that learns a new 'word', an embedding, for a concept without changing the model's weights.
The AI look: The polished, oversmooth, conventionally pretty aesthetic that marks an image as machine-made, a result of models regressing toward the average of their training data.→ Black Forest Labs (FLUX)
Throughput: How much a system can generate over time, for example images per minute, as opposed to latency, the wait for a single result.→ The economics of generative imagery
Tiled diffusion: Generating or upscaling a large image in overlapping patches and stitching them, a way to exceed what a single pass can hold in memory.→ Upscaling models
Tiling: Splitting a large image into overlapping tiles, processing each, and stitching them back, a way to generate or upscale beyond what memory allows.
Timestep: A single point in the diffusion process, from fully noised to fully clean. Models are conditioned on the timestep so they know how much noise to expect.→ The deep history, part two: the transformer takes over
Token: A single unit of input, a word fragment for text or a patch for images, that a model processes.→ The deep history, part two: the transformer takes over
Tokenizer: The component that chops text or an image into discrete pieces called tokens, the units a model actually reads.→ Alibaba (Qwen-Image, Tongyi Wanxiang)
Training: Teaching a model from scratch by showing it many examples and updating its parameters. The slow, expensive part of the model lifecycle.→ Orientation: the one-paragraph version, and the one-page version
Transformer: The neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks.→ What is a generative visual model, really?
Trigger word: A unique, deliberately unusual token used in training captions to name what a LoRA teaches, so that including the word in a prompt activates the learned subject or style.→ The training dataset is where LoRAs are made or broken
Turbo / Lightning: Labels for distilled, few-step versions of image models that trade a little quality for a large speedup.
Two-consents rule: The standard that one consent is needed to create a digital replica of a performer and a separate consent is needed for each use of it.→ The reckoning: licensing, likeness, and consent

U

U-Net: A convolutional neural network architecture with a characteristic U shape (compress, then expand) used as the backbone of early diffusion models. Largely replaced by Diffusion Transformers in 2023-2024.→ The deep history, part two: the transformer takes over
Uncanny valley: The dip into discomfort that happens when an artificial human is realistic enough to read as human but wrong in small details, named by Masahiro Mori in 1970.→ Questions readers always ask
Unit economics: The per-unit costs and revenue of a product, for example the cost to generate one image against the price charged for it.→ The economics of generative imagery
Upscaler: A model or tool that raises the resolution of an image or video, either faithfully recovering detail or inventing plausible new detail.→ Upscaling models
Upscaling: Increasing an image or video's resolution after generation, often with a model that invents plausible fine detail rather than simply stretching pixels.→ The single most important piece of infrastructure: ComfyUI

V

VAE (Variational Autoencoder): A neural network that learns to compress images into a small numerical representation (a latent) and decompress them back. Used in latent diffusion models to make the generation process tractable.→ The deep history, part one: from 2015 to Stable Diffusion
VALL-E: A Microsoft speech model that treats text-to-speech as predicting audio codec tokens with a language model, cloning a voice from a roughly three second sample. Its lineage defines the token-prediction approach to speech.→ How audio generation fits, in brief
VFX (Visual Effects): Imagery created or altered beyond what was captured on camera, now increasingly generated.→ Runway (Gen-3, Gen-4, Aleph, Act-Two)
Vibe directing: Directing a generative piece by intent and reference rather than by hand, describing the feel you want and steering it through iterations.→ Vibe directing: the new posture
Video-to-audio: Generating a soundtrack timed to an existing video clip, so effects land on the exact frame of the on-screen action. Also called video-to-sound.→ The music and sound landscape
Virtual production: Shooting live action against LED walls or in-camera digital backgrounds instead of a green screen, often using AI-generated environments.→ 3D generation models
Virtual try-on (VTO): Placing a garment or product onto a model or person in an image, so a catalog can be produced without a physical photoshoot.→ Inside fal: one platform up close
Vision Transformer (ViT): A transformer applied to images by splitting a picture into patches and treating each patch as a token, the idea that let attention-based models take over computer vision.→ The deep history, part two: the transformer takes over
Voice cloning: Reproducing a specific person's voice from a short sample so it can speak new lines, often in any language.→ How audio generation actually works
Voicebox: Meta's text-to-speech model built on flow matching, which sharpens a whole clip out of noise rather than predicting tokens one by one.→ How audio generation fits, in brief
VRAM (Video RAM): The memory on a GPU. It sets a hard ceiling on the model size and image resolution you can run locally.→ Black Forest Labs (FLUX)

W

Walled garden: A platform where content can be created, customized, and shared internally but never exported, the position Udio took after its licensing deals.→ Udio
Watermarking: Hiding a detectable signal inside generated media so its origin can later be identified. For audio and AI media it remains an arms race rather than a guarantee.→ How audio generation actually works
Wav2Lip: The 2020 open-source lip-sync model, trained against a dedicated lip-sync expert, that remains widely deployed and seeded the commercial lip-sync tools that followed.→ Lipsync and audio-visual synchronization
Weights: The learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable.→ What is a generative visual model, really?
World model: A model that learns an internal simulation of how a scene evolves, aiming to keep physics, objects, and space consistent over time. A frontier for long, coherent video.→ The landmark papers, and why they mattered

Z

Zero-shot voice cloning: Reproducing a specific voice from just a few seconds of reference audio with no per-speaker training, because the model reads the sample in context and then speaks new lines in that voice.→ The audio frontier: speech, music, and sound catch up