AI Glossary

A way to represent a scene as millions of tiny colored blobs (Gaussians) instead of a triangle Mesh. Fast to render, great for turntables and novel views from photos. Different from NeRF and from Hunyuan3D-style generated meshes.

A community-trained upscaler model highly prized for its ability to increase resolution without adding the "plastic" smoothing effect of other models.

An open music generation model (ACE-Step 1.5 and related) for local text-to-song and continuation. An alternative to closed Suno / Udio when you want weights, LoRA, or ComfyUI graphs for music.

After Detailer; an extension that automatically detects faces or hands in a generated image and runs a second inpainting pass to fix distortions.

AI Generated Content; a broad term referring to any media created by generative artificial intelligence.

A ComfyUI node that calls a cloud model (fal.ai, Replicate, a lab API) instead of a local Checkpoint. Mixes local ControlNet/upscale with closed models like Kling or Veo in one Workflow. Needs an API key and sends your Prompt off-machine.

A loop where an LLM calls tools, reads results, and continues until a goal is met. In creative stacks that can mean auto-building workflows, batch captioning, or a World Model agent that explores a generated scene.

AYS — a Scheduler / timestep recipe that picks better Step positions so 10–20 step renders look closer to 30–50. Used with SDXL and some video Samplers when you want speed without switching to a distilled Checkpoint.

A property of certain samplers (marked with 'a', like Euler a) that adds fresh noise at each step, resulting in non-convergent, constantly shifting images.

A type of sampler (like Euler a) that adds random noise during the generation steps, resulting in non-convergent, ever-changing images.

A motion module that can be attached to Stable Diffusion models to generate consistent, high-quality short video clips from text prompts.

An upscaling model specifically trained on anime line art and shading, preventing the over-sharpening of 2D illustrations.

Unwanted visual errors in an AI image, such as extra fingers, distorted limbs, or pixelated noise.

A grayscale image used during inpainting that tells the model exactly which pixels to change (white) and which to keep frozen (black).

A technique that allows neural networks to focus on specific parts of input data (like important words in a prompt) while ignoring less relevant details.

An open flow-matching image model (Fal / community) that sits beside Flux and SD 3.5. Uses Rectified Flow rather than classic epsilon diffusion. Needs its own Text Encoder stack and Sampler settings in ComfyUI.

The "AE" in VAE; a neural network trained to copy its input to its output, used to compress images into Latent Space.

A popular open-source web interface (WebUI) for running Stable Diffusion locally with a vast ecosystem of extensions.

Bfloat16, a 16-bit format used for training and Inference on modern NVIDIA GPUs. More stable than FP16 for large Transformers. Many Flux, Qwen-Image, and video checkpoints ship as BF16 safetensors.

Bootstrapping Language-Image Pre-training; an AI tool used to automatically generate descriptive captions for training images.

A syntax command in Automatic1111 that forces the text encoder to process the prompt in separate chunks to prevent concepts from bleeding into each other.

Isolating a subject onto alpha. RMBG, BiRefNet, and SAM-based nodes do this in ComfyUI. First step for stickers, product shots, and compositing a character onto a new scene before Outpainting the rest.

The central algorithm in training neural networks where error rates are calculated and sent backward to adjust weights.

The number of images generated simultaneously in a single operation.

The tendency of an AI model to produce results that favor certain demographics, styles, or concepts due to skewed training data.

A prompting style that uses specific tags (e.g., "1girl, blue_hair, solo") separated by commas, popularized by anime models like Pony Diffusion.

A training technique where images of different aspect ratios are grouped together so they don't have to be cropped or distorted during the learning process.

A setting in Dynamic Thresholding that attempts to make high CFG scales (which usually burn images) look as natural as low CFG scales.

Classifier Free Guidance scale; a parameter that determines how strictly the model adheres to your text prompt versus its own creative freedom.

Contrastive Language-Image Pretraining; the neural network used by Stable Diffusion to understand the relationship between text and images.

A setting that determines how early the process of text interpretation is stopped; increasing it can make prompts less literal and more abstract.

The "BigG" text encoder used in SDXL and SD3; it understands complex sentence structures better than the simpler CLIP-L.

The standard text encoder used in SD 1.5; it focuses on specific keywords and tags rather than natural language sentences.

Compute Unified Device Architecture; a parallel computing platform by NVIDIA that allows AI models to run efficiently on GPUs.

A ControlNet preprocessor that creates a line-art map of an image's edges to guide the generation of a new image with the same structure.

Keeping the same person or creature across shots, styles, and clips. Tools: a character LoRA, InstantID / PuLID, Flux Kontext references, Nano Banana chat memory, or First-Last Frame for video. The unsolved core problem of AI series work.

A fully trained model file (usually .ckpt or .safetensors) that contains all the weights and knowledge required to generate images.

An open image model line in the Flux/flow family, often used as a lighter or more permissive alternative Checkpoint in ComfyUI. Check the license and which Text Encoder it expects before swapping it into a Flux workflow.

The largest community platform for hosting, sharing, and downloading Stable Diffusion models, LoRAs, and embeddings.

The method behind CFG Scale. The model runs a conditioned pass (your Prompt) and an unconditioned pass, then steers toward the prompt. Higher CFG follows the text more strictly; too high burns or oversaturates the image.

A tool that analyzes an image and attempts to reverse-engineer a text prompt that could generate a similar result.

A node-based graphical interface for Stable Diffusion that allows users to build complex, modular image generation workflows.

The ComfyUI extension that installs, updates, and disables Custom Node packs from git. First thing to add on a new install. Missing nodes in a shared Workflow almost always mean Manager has not pulled that pack yet.

The extra tensors that steer a diffusion or flow model: text embeddings from CLIP or T5 Encoder, ControlNet maps, IP-Adapter image features, or audio. Positive and negative Conditioning are combined with CFG Scale.

An alternative to the standard VAE decoder developed by OpenAI that attempts to produce sharper, less glitchy images from latent space.

How much Token history a Transformer can attend to at once. In LLMs it is conversation length; in DiT image/video models it is how much prompt, reference, or frame context fits. Longer windows need more VRAM.

An advanced neural network extension that allows users to control diffusion models using conditions like edge detection, depth maps, or pose estimation.

The point in training or generation where the model settles on a final output and further steps yield diminishing changes.

The mechanism where the model connects visual data (pixels) with semantic data (text prompts) to generate relevant images.

A community ComfyUI plugin (Kijai wrappers, rgthree, Impact Pack, GGUF loaders). Installed via ComfyUI Manager or git clone into custom_nodes. Broken graphs after an update are usually a Custom Node version mismatch.

Denoising Diffusion Implicit Models; one of the first and fastest samplers, known for being deterministic (same seed = same image) but requiring more steps for detail.

A sampler that ignores the user's step count and automatically determines how many steps are needed to reach a clean image.

A second-order multistep sampler that is extremely fast and balances quality with speed; widely considered the "default" for modern generation.

An ancestral version of the DPM++ sampler; slower than 2M but produces softer, more painterly results due to added noise.

A third-order solver using stochastic differential equations; very slow but produces the highest theoretical accuracy for complex textures.

A sophisticated sampler known for producing high-quality, realistic details at lower step counts using stochastic differential equations.

A whole-body pose estimator (body, hands, face) used as ControlNet / OpenPose input. More reliable fingers and faces than older OpenPose in ComfyUI, especially for dance and Image-to-Video pose follow.

A structured collection of data (images and text) used to train or fine-tune an AI model.

An advanced extension for creating frame-by-frame animations, often featuring 3D camera movement and trippy, morphing transitions.

A setting in Image-to-Image workflows that controls how much the AI is allowed to change the original input image.

A grayscale representation of an image where lightness indicates proximity to the camera; often used with ControlNet to preserve 3D composition.

Diffusion Transformer — a diffusion backbone that uses a Transformer instead of a U-Net. Flux, SD 3, Hunyuan Video, and most 2025–2026 image/video models are DiTs, which scale better than U-Net at high resolution.

A type of generative model that learns to create images by gradually removing noise from a random static signal.

Training a smaller or faster student model to copy a larger teacher. Distilled checkpoints (Turbo, Lightning, LCM, Schnell, many video accelerators) generate in far fewer Steps, usually with a narrower CFG Scale range.

The version of CFG Scale used by newer models like Flux.1; it controls how much the model separates the image from the noise.

Weight-Decomposed Low-Rank Adaptation. Splits each weight into magnitude and direction before applying a LoRA-style update. Often matches or beats LoRA quality at the same rank, and is supported in Kohya_ss and several ComfyUI trainers.

A training technique used to fine-tune a model on a specific subject, person, or style using a small set of reference images.

Using two Text Encoder towers at once, as in SDXL (CLIP-L + CLIP-G) or Flux (CLIP-L + T5 Encoder). Each encoder sees the Prompt; their embeddings are concatenated or pooled before the backbone.

A technique that allows you to use extremely high CFG Scales (like 20 or 30) without color-burning the image, useful for strict prompt adherence.

Exponential Moving Average; a version of model weights saved during training that is smoother and more stable than the raw weights.

Enhanced Super-Resolution Generative Adversarial Networks; a popular AI upscaling model used to sharpen and enlarge images.

Enhanced Super-Resolution GAN; an older but reliable upscaler that sharpens images by hallucinating realistic textures onto low-res inputs.

Also known as Textual Inversion; a small file that points to a specific concept or style within a model's existing knowledge.

The act of combining multiple Textual Inversion files into one, though this often dilutes the effectiveness of each.

One complete pass through the entire training dataset during the model training process.

The standard Euler method; a first-order solver that is the simplest and fastest way to solve diffusion equations, though less accurate than Heun.

A popular ancestral sampler known for speed and creativity, though the results change if the step count is altered.

Half-Precision Floating Point; a data format that uses 16 bits instead of 32, reducing VRAM usage and speeding up generation with minimal quality loss.

Single-Precision Floating Point; the standard 32-bit data format offering high accuracy but requiring significant memory.

8-bit floating point weights. Compared with FP16 or BF16, FP8 roughly halves VRAM for the same U-Net or DiT, with a small quality tradeoff. Common for Flux, Wan, and other large models on 12–16 GB cards.

Frames per second. Many local video models generate 12–16 FPS and rely on Frame Interpolation to reach 24 FPS film cadence. Generating native 24 FPS costs more VRAM and time; upscaling FPS after the fact is the usual compromise.

Replacing a face in an image or video with another identity. Reactor is the usual ComfyUI node; InsightFace provides the embeddings. Brand-safety and consent matter — this is the same tech as Deepfakes.

The process of taking a pre-trained model and training it further on a specific dataset to specialize it.

A video Control where you supply the opening and ending stills and the model interpolates the shot between them. Gives much more narrative control than Text-to-Video alone; supported in many Wan, Kling, and Runway-style workflows.

A fast, memory-efficient Attention Mechanism implementation. Cuts VRAM and speeds Transformer Inference on NVIDIA GPUs. Sage Attention is a later alternative used a lot in ComfyUI video graphs.

A Microsoft VLM used as a Dataset captioner and as a ComfyUI interrogator. Produces natural-language captions (not Booru tags). A common first step before LoRA training, alongside JoyCaption and WD14 Tagger.

The advanced generation architecture used by Flux and SD3 that creates a straighter, more efficient path from noise to image than standard Diffusion.

A highly advanced open-source image generation model released in 2024, known for superior prompt adherence and text rendering compared to SDXL.

Black Forest Labs' follow-up to the original Flux family, with stronger editing, typography, and reference handling than Flux Dev. Local GGUF and FP8 builds are used in ComfyUI alongside Kontext-style edit workflows.

The full-quality open-weight Flux image model from Black Forest Labs. Strong prompt adherence and text rendering. Heavier than Flux Schnell; typically run in FP8 or GGUF on 12–24 GB GPUs, often with a LoRA for style.

Flux instruction/edit models that take an image plus a natural-language edit ("make it night", "change the outfit") instead of a full Txt2Img Prompt. Strong at Character Consistency and local edits without a separate Inpainting Model.

The distilled, Apache-licensed Flux variant. Generates in about 4 Steps with a low CFG Scale. Faster and more permissive than Flux Dev, with slightly weaker prompt following — the usual choice for previews and bulk renders.

A simplified local image UI (based on Gradio + Comfy internals) that hides most Samplers and models. Strong defaults, inpaint, and image prompt. Less flexible than ComfyUI, much faster to learn.

A high-performance Automatic1111 fork with better VRAM hacks and Unet patching. Many SDXL and some Flux extensions landed here first. reForge is a community continuation after upstream Forge slowed down.

Inventing in-between frames to raise FPS (for example 16 → 24 or 24 → 48). RIFE and similar networks do this after generation. Fixes choppy AI video but can smear fast motion if the source frames already conflict.

A backbone tweak that reweights skip connections in a U-Net at Inference time — no extra training. Can boost sharpness and cohesion on SD 1.5 / SDXL. Not a drop-in for DiT models like Flux.

Generative Adversarial Network; an older architecture where two neural networks (a generator and a discriminator) compete to create realistic data.

GPT-Generated Unified Format; a highly compressed file format originally for LLMs, now used to run massive image models like Flux on lower VRAM GPUs.

Storing a U-Net or DiT as a llama.cpp-style GGUF file (Q4, Q5, Q8). Lets Flux, Wan, and Qwen-Image run on 8–12 GB VRAM. Quality climbs Q4 → Q8; VRAM climbs with it. Needs a GGUF loader Custom Node, not a normal Checkpoint loader.

Graphics Processing Unit; the hardware component essential for running diffusion models and training neural networks quickly.

A subset of artificial intelligence focused on creating new content, including images, text, and audio, rather than just analyzing existing data.

A training technique that splits a batch of data into smaller mini-batches to allow training on GPUs with lower VRAM.

An optimization algorithm used during training to minimize errors by iteratively moving towards the lowest error value.

xAI's image and video generator. Known for fast iterations, fewer content filters than many closed labs, and shareable image/video posts. Closed model — you use it through grok.com, not a local Safetensors file.

A parameter in Flux models that acts as a secondary CFG scale, specifically influencing contrast and saturation.

A phenomenon where an AI generates confident but factually incorrect or nonsensical information; in images, this appears as visual glitches.

A sampler that acts as an improved version of Euler; it runs at half the speed (2 evaluations per step) but is much more accurate and cleaner.

An open DiT image model that appeared as a Flux alternative, with large text encoders and strong prompt following. Local ComfyUI support via FP8 and quantized builds; heavier than SDXL, similar class to Flux Dev.

An image processing technique used in upscaling to retain sharp edges while ignoring low-frequency color gradients.

High Resolution Fix; a workflow that generates an image at a lower resolution first and then upscales it to avoid distortion.

A major platform and community hub for hosting, sharing, and collaborating on open-source machine learning models and datasets.

Tencent's open video model. High quality Text-to-Video and Image-to-Video with large VRAM demand unless quantized. The same lab's Hunyuan3D line covers meshes; do not confuse the two checkpoints.

Tencent's open text/image-to-3D family (meshes, textures, later PBR and world tools such as WorldClaw). Outputs a Mesh you can drop into Blender, Unreal, or a game engine, unlike video models that only output pixels.

An older method of fine-tuning models that modifies the weights of the cross-attention layers; largely replaced by LoRA.

A configuration setting external to the model (like learning rate or batch size) set by the user before training begins.

Image Prompt Adapter; a powerful tool that allows you to use an image as a prompt to transfer style or composition without complex training.

A closed image model that specialized early in readable text inside images — logos, posters, comics. Still a common pick when Flux or SDXL struggle with long on-image lettering.

A major SDXL anime Checkpoint family trained on heavily tagged Booru Style data. Understands Danbooru-style tags, character names, and artist tags better than base SDXL. The usual base for modern anime LoRA on Civitai.

A generation mode where an existing image is used as the starting point instead of pure random noise.

I2V — animating a still image into a clip. The image locks identity and composition; the model adds motion. Usually more controllable than Text-to-Video and the default path for character sheets, keyframes, and product shots.

The process of using a trained model to generate new outputs (e.g., creating an image from a prompt).

The time it takes for a trained model to process an input and generate an output, usually measured in iterations per second (it/s).

The process of regenerating only a specific masked area of an image while keeping the rest unchanged.

A specialized checkpoint trained specifically to understand masked areas, performing better at filling gaps than standard models.

A face-identity method that uses InsightFace embeddings plus ControlNet-style conditioning so a person survives style changes. Heavier than a tiny FaceID LoRA, lighter than Dreambooth. Common in ComfyUI with IP-Adapter.

The repeating cycle of refining an image step-by-step from noise to clarity.

A structured Prompt as key/value fields (subject, camera, lighting, negative, duration) instead of one prose paragraph. Easier for Agents and for models that were instruction-tuned on JSON. Popular with Gemini, Grok, and Qwen image stacks.

An open captioning VLM fine-tuned for rich, accurate image descriptions. Preferred by many LoRA trainers over BLIP or basic Clip Interrogator because captions read like a real Prompt instead of a tag salad.

A library of sampling schedulers (like Euler, LMS, Heun) implemented by Katherine Crowson, widely used in Stable Diffusion.

A specific noise schedule (sigma) proposed by Nvidia researcher Tero Karras that improves the quality of samplers like DPM++ at lower steps.

Kuaishou's closed video model. Strong cinematic motion, camera moves, and later omni/audio versions (including Kling O1). API/web only — a common cloud alternative when local Wan or LTX is not enough.

A popular GUI and script collection used for training LoRAs and fine-tuning Stable Diffusion models.

A real-time and enhancement-focused image platform and model line. Krea 2 is used for high-detail generation and latent upscale workflows; community ComfyUI graphs often pair it with custom Upscaler or LSR nodes.

Latent Consistency Model; a distillation technology that allows models to generate high-quality images in as few as 4 to 8 steps.

Large Language Model; an AI trained on vast amounts of text data to understand and generate human language (e.g., GPT-4, Claude).

Linear Multi-Step; a sampler that uses data from previous steps to predict the next step, offering a smoother convergence than Euler.

Lightricks' open video models (LTX Video, LTX-2, LTX 2.5). Fast latent video generation, including open weights and audio-aware later versions. Popular in ComfyUI when you want quicker previews than full Wan or Hunyuan Video.

A traditional non-AI resampling algorithm used for downscaling images or simple upscaling where preserving exact pixel values is preferred over smoothing.

The specific architecture of Stable Diffusion where the diffusion process happens in a compressed latent space rather than pixel space.

A compressed mathematical representation of data where the diffusion model actually operates before decoding the result into pixels.

A method of upscaling the image while it is still in the compressed Latent Space, which is faster but can introduce blurry artifacts.

A hyperparameter that controls how much the model's weights are adjusted during each step of training; too high causes instability.

A setting that changes the speed of learning during training (e.g., Constant, Cosine), allowing for fine adjustments as the model gets smarter.

Driving a face so mouth shapes match a speech or song track. Used with talking-head Image-to-Video, InfiniteTalk-style models, and audio-to-video pipelines. Needs a clean audio clip and a stable identity frame.

Low-Rank Adaptation; a lightweight training technique that adds specific characters, styles, or concepts to a model without modifying the large base checkpoint.

The size of a LoRA adapter (network dimension). Rank 8–16 is typical for a style; 32–128 is common for a character or concept. Higher rank captures more detail, uses more VRAM, and overfits faster if the Dataset is small.

A variation of LoRA that targets different layers of the neural network, often capable of learning more complex styles or motions.

An open autoregressive / diffusion-forcing video model line aimed at longer, more coherent clips than a single-shot latent video model. Used in ComfyUI for extended Image-to-Video with history or teacher-forcing style context.

Model Context Protocol — a standard way for AI apps to expose tools (files, browsers, ComfyUI, game engines) to an Agent. Lets a chat model drive local software without a one-off plugin for every host.

Multimodal Diffusion Transformer; the architecture behind Stable Diffusion 3 that processes text and images in separate weights to improve text adherence.

A 3D surface made of vertices and triangles (OBJ, GLB, FBX). Image models output pixels; 3D models output a Mesh you can rig, UV, and texture. Topology quality decides whether it is usable in a game or only as a sculpt preview.

A closed Discord/web image model famous for tasteful default aesthetics. Versions through V7 added better prompt following, draft mode, and omni-reference. Not run locally; parameters like --ar, --stylize, and --oref are Midjourney-specific.

The lab behind Hailuo video and MiniMax music models (including Hailuo / MiniMax H3 video and music generators). Closed or partially open depending on the release; known for realistic people and longer clips than early open video models.

A failure in GAN training where the generator starts producing the same limited set of outputs regardless of the input.

A short hexadecimal code (e.g., "1234abcd") used to uniquely identify a specific version of a model file to prevent confusion.

The process of combining the weights of two or more checkpoints to create a hybrid model with characteristics of both.

The process of removing unnecessary data (like EMA weights) from a model file to reduce its size (e.g., FP16) without losing quality.

A LoRA trained on camera or body motion rather than appearance — orbit, zoom, walk cycle, dance. Applied on top of AnimateDiff, Wan, or Hunyuan Video to reuse a move without restating it in the Prompt every time.

Generating several camera angles of the same subject (front, side, back, 3/4) that share identity. Used as input to Text-to-3D, character sheets, and Qwen multi-angle LoRA. Inconsistent views are the main failure mode.

Also known as Tiled Diffusion; a technique that allows the generation of massive images (e.g., 4k or 8k) by processing them in small grid tiles to save VRAM.

A model that takes more than one kind of input — text, image, video, audio, or pose — in the same network. VLMs, Flux Kontext, Qwen-Image-Edit, and most 2026 video models are multimodal rather than text-only.

Normal Float 4-bit; an extreme quantization format that allows massive models (like Flux Dev) to fit into consumer graphics cards with minimal quality loss.

Community name for Google Gemini's image generation/editing path. Strong at Character Consistency across a conversation and at following natural-language edits. Closed API, not a local Checkpoint.

Neural Radiance Field — a neural scene that you can re-render from new camera angles. Predecessor to 3D Gaussian Splatting for photogrammetry-like capture. Heavier to train and slower to render than splats.

The simplest upscaling method that just repeats pixels; useful for pixel art but produces jagged "staircase" edges on photos.

A text input field where you define concepts, objects, or styles you want the AI to avoid generating.

A computing system inspired by biological brains, consisting of interconnected nodes (neurons) that process information in layers.

Random pixel data (static) that acts as the starting canvas for a diffusion model.

A mathematical curve that dictates how much noise is removed at each step of the generation process (e.g., Karras, Exponential, Uniform).

An Illustrious-line anime Checkpoint (often called NAIV or NoobAI-XL) with a huge tag vocabulary. Favored for character accuracy and for training new character LoRA. Prompting is tag-based, not only natural language.

An SVDQuant-based quantized inference engine for Flux-class DiTs in ComfyUI. Aims for 4-bit quality closer to BF16 than naive NF4. An alternative to GGUF when you want speed on NVIDIA without a full FP8 model.

Trying to get the perfect result in a single generation without using inpainting, editing, or refiners.

A model whose trained Weights are downloadable for local Inference, unlike a closed API. Licenses still vary — Apache 2.0, MIT, OpenRAIL, or non-commercial. Flux Dev, Wan, LTX, and Qwen-Image are open-weight; Sora and Veo are not.

A ControlNet model that detects human stick-figure skeletons to force the AI to generate a character in a specific pose.

A map of where pixels move from frame to frame. Used inside Frame Interpolation, video-to-video warping, and some Temporal Consistency tricks. Bad flow (fast pans, explosions) is why interpolators smear.

A mathematical framework used in newer models (like Flux) to find the most efficient path between noise and data.

The process of extending an image beyond its original borders by generating new consistent content.

A slang term for training a LoRA or model for too many epochs, causing it to lose flexibility and produce fried, high-contrast artifacts.

Also known as Overfitting; training a model so long that it memorizes the data and loses creativity.

A training failure where a model learns the training data too well, resulting in an inability to generate new, creative variations.

Perturbed Attention Guidance — an Inference trick that perturbs self-attention instead of raising CFG Scale. Can add detail and reduce burnt looks on SDXL. Implemented as a ComfyUI node; not the same as PAG in LLMs.

Physically Based Rendering textures: albedo, roughness, metallic, normal (and often AO/height). A PBR-ready Mesh lights correctly in Blender, Unreal, or Unity. Hunyuan3D, Chord, and HiTem3D-style tools aim to emit these maps, not just a baked color PNG.

Parameter-Efficient Fine-Tuning. Umbrella for LoRA, QLoRA, DoRA, LyCORIS, and similar methods that train a small adapter instead of every Weight in the Checkpoint.

Pseudo Linear Multi-Step; an older sampler originally used in Stable Diffusion 1.4, largely obsolete now due to the superior DPM++ family.

A configurable variable in a model or workflow, such as steps, seed, or CFG scale.

The total number of adjustable weights in a model (e.g., SDXL has 6.6 billion), indicating its complexity and capacity.

A one-click launcher for local AI apps — ComfyUI, Forge, video packs, music models — with install scripts and isolated environments. Aimed at people who do not want to manage Python venvs by hand.

A highly influential SDXL-based model ecosystem known for its strict adherence to "Booru Style" tags and high aesthetic quality.

The text description input by a user to guide the AI in generating a specific output.

A syntax technique (e.g., [cat:dog:0.5]) that tells the AI to generate a cat for the first 50% of steps, then switch to generating a dog.

The art and science of crafting precise text inputs to guide an AI model toward a specific desired result.

Search and Replace; a script in Automatic1111 that allows you to generate a matrix of images by swapping out one word (e.g., "cat", "dog", "bird") in the prompt.

The ability to activate or deactivate specific parts of a prompt at specific step numbers during generation.

The syntax (usually (word:1.2) or (word:0.8)) used to tell the AI to pay more or less attention to a specific token.

A model file that has had unnecessary training data (like EMA weights) removed to reduce file size for inference.

Pure and Lightning ID — a high-fidelity identity injector for Txt2Img and Image-to-Image. Often sharper likeness than older IP-Adapter FaceID at the cost of extra nodes. Used when InstantID is too plastic or too weak.

An open-source machine learning framework developed by Meta, widely used for building and training deep learning models like Stable Diffusion.

Quantized LoRA training: the base Checkpoint is loaded in 4-bit (often NF4) while LoRA adapters train in higher precision. Lets you Fine-tune Flux-class models on consumer VRAM that could not hold a full BF16 train.

The process of reducing the precision of a model's weights (e.g., from 16-bit to 8-bit or 4-bit) to drastically lower VRAM requirements.

Alibaba's open image stack built on the Qwen VLM family. Known for bilingual prompts, text in-image, and a strong edit variant (Qwen-Image-Edit). Competes with Flux for local Txt2Img and instruction editing.

Real-ESRGAN; an updated version of ESRGAN optimized for photographic images, fixing the "blobby" artifacts of the original.

Retrieval-Augmented Generation. The model searches your docs, glossary, or prompt library, then answers using those chunks. Keeps an LLM current without Fine-tuning, and reduces Hallucination on site-specific facts.

Real-Time Intermediate Flow Estimation — a widely used Frame Interpolation network. A standard ComfyUI node for smoothing 12–16 FPS AI clips up to 24 or 32 FPS before upload.

Reinforcement Learning from Human Feedback. After supervised Fine-tuning, a reward model trained on human rankings is used to push the LLM toward preferred answers. How most chat models learned to be helpful and refuse some content.

A fast and accurate face-swap extension for Stable Diffusion that replaced the discontinued Roop project.

An LLM trained to think in hidden chain-of-thought before answering — useful for planning ComfyUI graphs, debugging errors, or writing structured JSON Prompt blocks. Slower and more expensive per Token than a chat model.

A newer generation technique used in models like Flux, which creates a straighter, more efficient path from noise to image than standard diffusion.

A still you feed as Conditioning — style, pose, identity, or composition — instead of describing everything in the Prompt. IP-Adapter, Flux Redux/Kontext, ControlNet, and most Image-to-Video pipelines start from a reference.

A ControlNet mode that uses a reference image to guide the style or content of generation without using weights.

A secondary model used in workflows (like SDXL) to polish the image, add details, and remove noise after the initial generation.

An extension that allows you to split your image into distinct sections (e.g., left, right, top-left) and apply different prompts to each area.

A set of generic images (e.g., photos of random dogs) used during Dreambooth training to stop the model from forgetting what a "dog" is while learning a specific pet.

An upscaler model specialized in retaining fine details in textures like fabric, skin, and landscapes without over-smoothing.

The final act of generating the visual output.

A cloud Inference host where models run from a web API or Cog container. Similar role to fal.ai: rent a GPU for Flux, video, or 3D jobs without a local 24 GB card.

A unique sampler that pauses during generation, adds noise back in (restarts), and denoises again to refine details deeper than standard solvers.

Stable Diffusion v1.5; a legendary base model known for its speed and massive library of community-created resources.

The upgraded version of Stable Diffusion 3, featuring improved prompt adherence and fixing the "body horror" issues of the initial release.

Stable Diffusion XL; a significant evolution of Stable Diffusion with a larger architecture and dual text encoders for higher resolution and better comprehension.

A distilled version of SDXL capable of generating high-quality images in just 1 step, allowing for real-time generation.

Supervised Fine-Tuning — training on prompt/response pairs. The first stage after pretraining for chat models, captioners, and many instruction-tuned image editors such as Flux Kontext and Qwen-Image-Edit.

A restoration Upscaler that uses a diffusion prior to add realistic detail, not just sharpness. Heavier than ESRGAN or 4x-UltraSharp; used when you want a 1K AI still to survive a 4K crop without looking like plastic.

A secure file format for saving AI models that prevents the execution of malicious code, unlike the older .ckpt format.

A filter built into many diffusers pipelines that scans the generated image for NSFW content and blurs it to black if detected.

An advanced attention mechanism optimization that significantly improves processing speed and reduces memory usage during inference.

The algorithm (e.g., Euler, DPM++, K-LMS) that calculates how to remove noise from the image at each step.

The noise timetable a Sampler follows — how much noise is removed at each Step. Karras Schedule, simple, beta, and exponential are common. Sampler and scheduler together decide speed, sharpness, and how many steps you actually need.

A numerical value used to initialize the random noise generator; using the same seed and settings will result in an identical image.

A video restoration / Upscaler model used to sharpen and denoise AI clips (often after Wan or LTX). Runs as tiled video upscale in ComfyUI; slower than RIFE but recovers detail instead of only inventing frames.

SAM — a foundation model that masks any object you click or box. In ComfyUI it drives inpaint regions, background removal, and regional Prompt. Grounding DINO is often paired with it for text-to-mask.

A mechanism where the model looks at different parts of the image itself to ensure consistency (e.g., making sure two eyes look the same).

A setting in Flux schedulers that shifts the noise focus; higher values improve high-resolution details but may require more steps.

A Google contrastive vision-language encoder, successor-style to CLIP in many 2025+ models. Used as a vision tower for captioning, IP-Adapter-like reference, and some DiT conditioners. Not a drop-in CLIP-L replacement.

Motion stored as joint transforms on a rig, not as baked vertices. Needed for games and cinematics. Text-to-motion tools (kimodo.cpp, MotionBricks-style systems) output skeletons you attach to a Mesh.

A technique to blend the inpainted area seamlessly with the surrounding pixels by softening the mask edges.

OpenAI's closed Text-to-Video / Image-to-Video system. Set the 2024 quality bar for physics and shot length. Not available as local Weights; most local workflows try to approximate it with Wan, Hunyuan Video, or LTX.

Stability AI's open-ish audio family for sound effects and music beds. More instrumental/SFX than full sung tracks compared with Suno, and runnable locally with the right Checkpoint and VRAM.

The number of iterations the model takes to refine the noise into a clear image; more steps usually mean higher quality up to a point.

The ability to apply the artistic style of one image to the content of another.

A nested ComfyUI graph packaged as one node. Cleans huge Workflows (Load Checkpoint + LoRA + Sampler) into a single block you can reuse. Breaking a shared JSON is often just an expanded subgraph the recipient does not have.

A closed text-to-song model. Strong vocals and full-track structure from a short Prompt. Not local; watch commercial terms and download limits. ACE-Step and Stable Audio are the usual open-weight alternatives.

A modern local UI over ComfyUI/backend workers, with a friendlier generate tab than raw graphs. Useful when you want Automatic1111-like simplicity but still need ComfyUI Custom Node power underneath.

Swin Transformer for Image Restoration; a high-end upscaler that uses transformer attention mechanisms for state-of-the-art restoration quality.

Hidden instructions that set an LLM or multimodal model's role, style, and limits before the user Prompt. In image tools it may live as a style preset or a fixed prefix. Changing it often matters more than tweaking the last user line.

Text-to-Image Adapter; similar to ControlNet, these are lightweight models that provide additional control guidance to the generation.

Text-To-Text Transfer Transformer; a massive text encoder used in Flux and SD3 that allows the model to understand complex text and render legible words on signs.

The large T5 Encoder used by Flux, SD 3.5, and several video models to read long Prompts. Often 4-bit or FP8 quantized. If text rendering or prompt following collapses after a model swap, the T5 file is usually missing or mismatched.

Microsoft's open structured 3D generator. Produces assets with cleaner geometry than many early image-to-3D hacks. Used alongside Hunyuan3D, Tripo, and Meshy in local and API 3D pipelines.

A caching trick for DiT Inference that skips recomputing early transformer blocks when the residual barely changes. Can cut Flux and Wan runtime a lot, with a small risk of blur or identity drift at aggressive settings.

A sampling control for LLMs and some multimodal models. Low temperature is deterministic and literal; high temperature is more random and creative. Image diffusion uses Seed and CFG Scale for a similar job.

How stable identity, lighting, and details stay across frames. Flicker, morphing faces, and melting backgrounds are consistency failures. Helped by Image-to-Video, Motion LoRA, overlapping context windows, and video-aware Upscalers.

A mathematical object (a multi-dimensional array) that creates the fundamental data structure used in neural networks.

An optimization engine by NVIDIA that compiles models specifically for your GPU, increasing generation speed by up to 40%.

The network that turns a Prompt into embeddings the U-Net or DiT can use. SD 1.5 uses CLIP-L; SDXL adds CLIP-G; Flux and SD 3.5 add a T5 Encoder. Swapping or skipping encoders changes prompt adherence and style.

Going from a Prompt (or a single image) to a Mesh plus textures. Hunyuan3D, TRELLIS, Tripo, and Meshy are common paths. Results still often need retopology and PBR cleanup before production use.

Generating audio from a Prompt: genre, mood, lyrics, or reference. Suno and Udio are closed; ACE-Step and Stable Audio are the local/open options. Separate from Lip Sync, which only matches an existing vocal to a face.

T2V — generating a clip from a Prompt only, with no starting frame. Wan, Hunyuan Video, LTX, Kling, Sora, and Veo all offer T2V. Harder than Txt2Img because motion, identity, and lighting must stay coherent across frames.

A technique for training a new word (token) into a model to represent a specific concept; see Embedding.

A VAE optimization that decodes the image in small chunks (tiles) to prevent Out of Memory (OOM) errors on large images.

A generation mode that forces the edges of an image to wrap around seamlessly, used for creating textures for 3D models or video games.

The basic unit of text that the CLIP model understands; a prompt is broken down into tokens before processing.

The maximum number of text tokens a model can process in a single prompt (often 75 or 77 in standard CLIP models).

ToMe; an optimization technique that speeds up generation by merging similar tokens in the attention layers, trading a tiny bit of quality for speed.

When an LLM emits a structured function call instead of plain text — search, generate an image, run a workflow, or read a file. The host executes the tool and feeds the result back. Core to Agent setups.

Nucleus sampling for LLMs: the model only samples from the smallest set of tokens whose probabilities sum to p (for example 0.9). Works with Temperature to stop the tail of nonsense tokens.

The massive collection of image-text pairs fed into the model during its initial creation to teach it what things look like.

A deep learning architecture originally built for text (like GPT) but now adapted for vision (Vision Transformers) in modern image models.

Text-to-Image; the standard workflow of generating an image from scratch using only a text prompt.

The core neural network architecture of Stable Diffusion responsible for predicting noise subtraction.

A closed music generator competing with Suno, with stem-like control and cover/extend features. Web/API only. Same caution as other commercial music models around training data and commercial use.

Unified Predictor-Corrector; a specialized sampler designed for extremely fast convergence, capable of good images in 10-15 steps.

The multiplier used during the High-Res Fix pass (e.g., 1.5x or 2.0x) to determine the final resolution.

An algorithm or model designed to increase the resolution of an image while attempting to add or preserve detail.

The process of increasing the resolution of an image, often using AI to fill in missing details.

Variational Autoencoder; the component responsible for compressing pixel images into latent space and decoding them back into viewable images.

A tiny, lightweight version of a VAE used solely to generate blurry, low-res preview images while the main image is still generating.

Vision-Language Model: an LLM that can see images (and often video). Used for captioning Datasets, Clip Interrogator-style reverse prompting, UI agents, and as the text/vision tower inside models like Qwen-Image.

Video Random Access Memory; the dedicated memory on a GPU crucial for loading large AI models and processing high-resolution images.

A feature in ComfyUI and Forge that automatically moves unused model parts from the GPU to the slower System RAM to prevent crashes.

A secondary seed used to create slight variations of an existing image (e.g., keeping the composition but changing the facial expression).

A numerical representation of a concept (word or image feature) within the model's high-dimensional latent space.

A compression technique used in VAEs to map continuous data into discrete finite values.

Google DeepMind's closed video generator (Veo 2, Veo 3, and later). Strong cinematic camera language and audio on newer versions. API/Gemini only, alongside Nano Banana for stills.

V2V — restyling or restabilizing an existing clip while keeping motion. Similar to Image-to-Image but across frames. Used for stylizing footage, fixing AI video, or applying a LoRA look to a whole sequence.

The part of the model responsible for "seeing" and understanding visual data, often used in Image-to-Image tasks.

A specific AI tool used to analyze anime-style images and generate accurate booru-style tags for training data.

Alibaba's open video DiT family (Wan 2.1, 2.2, 2.6, and related). Strong local Text-to-Video and Image-to-Video in ComfyUI, with GGUF and FP8 ports. A default choice when you want Sora-like motion without an API.

In prompting, the numerical importance assigned to a specific word (e.g., (cat:1.5)) to make the AI focus more or less on it.

The practice of adding emphasis to certain words in training data to teach the model that those concepts are more important.

The learnable parameters within a neural network that determine how input data is transformed into output.

A dynamic prompting technique where a placeholder (e.g., __color__) is replaced by a random word from a list during generation.

The specific sequence of nodes, settings, and models used to achieve a complex result (common in ComfyUI).

A network that predicts how a scene evolves if you act in it — look around, walk, interact — instead of emitting one clip. Used for explorable 3D/video environments and agent rollouts, distinct from a one-shot Text-to-Video render.

A script that generates a grid of images, changing two or three parameters (like CFG Scale vs Steps) to compare their effects visually.

A memory-efficient attention mechanism library developed by Meta that speeds up image generation and reduces VRAM usage.

In 3D computer graphics and ControlNet, the distance of a pixel from the camera, used to generate Depth Maps.

A fast distilled image model (Alibaba/Tongyi line) aimed at near-real-time local generation. Popular in ComfyUI for previews and high-volume jobs where Flux Dev would be too slow, often paired with an Upscaler afterward.

The ability of an AI model to complete a task (like generating a zebra in space) without having been explicitly trained on that specific scenario.

A serverless GPU API host for Flux, Wan, LTX, upscalers, and many open models. You pay per second instead of owning a GPU. Common when a Workflow is too big for local VRAM but you still want the same Checkpoint.

PyTorch 2 graph compiler. The first run is slow while it traces the model; later steps are faster. Popular in ComfyUI for Flux and video DiTs when you generate many images or frames with the same graph.