Short answer: this is the reference page. Every term is defined in two or three sentences, grouped by topic, with a link to the Lab article that goes deeper. If you meet a term in the wild that is not here, tell us and it will be.

Models and weights

Open weights. The model file is published and anyone can download and run it. This is the single most important property in adult AI: an open-weights model can be run without anyone’s content filter. Not the same as open source (the training data and code usually stay private).

Closed / API-only model. The model runs only on the vendor’s servers; you send requests and get results. Whatever the vendor’s policy blocks is blocked for everyone — see Wan 2.7 for how that plays out.

Checkpoint. A complete, ready-to-run set of model weights. In image generation, “checkpoint” usually means a community fine-tune of a base model — Lustify or CyberRealistic Pony are checkpoints of SDXL lineage.

Base model. The original model a family descends from (SDXL, Flux, Gemma 4). Capability and prompting style are inherited from it.

Fine-tune. A model further trained on new data to change behaviour: adding explicit anatomy to an image model, or roleplay ability to a chat model. Quality fine-tunes add skills; cheap ones just shift style.

LoRA. A small adapter file trained on top of a model that adds one concept — an act, a pose, a camera style, a face — without retraining the whole model. Stackable, swappable, and the reason Wan 2.2 is still in production. “Mods” on MyBabes are curated LoRAs.

Abliteration. A weight-level edit that removes a chat model’s refusal behaviour without adding any new skill. An abliterated model answers everything but roleplays no better than its base — the distinction matters, see Qwen 3.8 uncensored.

Quantisation. Compressing model weights to lower precision (Q8, Q4, FP8, NVFP4) so they fit on smaller GPUs, trading a little quality for a lot of memory. How a 22B video model fits a 24 GB card (LTX-2.5).

MoE (mixture of experts). A model built from many expert sub-networks with only a few active per token — big-model quality at smaller serving cost. The Serenity chat tune is one.

Distillation. Training a smaller, faster model to imitate a larger one. Distilled checkpoints are the “fast” variants you see everywhere.

Image generation

Diffusion model. The dominant image-generation architecture: starts from noise and denoises toward the prompt.

SD / SDXL / Pony / [Illustrious](/hub/models/illustrious). Stable Diffusion and its family tree. SDXL is the 2023 base; Pony XL and Illustrious are retrains of it with explicit content in the data. “Stable Diffusion porn” in 2026 means one of these three, not SD 1.5.

Tags vs natural language. Two prompting styles. Pony and Illustrious want comma-separated Danbooru-style tags; Flux and Krea want sentences. Using the wrong style is the most common cause of bad output.

Score tags. Pony’s quality dial: `score_9, score_8_up, score_7_up` at the start of a prompt pulls toward the best-rated training images.

Negative prompt. A second prompt listing what you don’t want (extra fingers, watermark, text). Essential on SDXL-family models, nearly unnecessary on Flux.

Seed. The random number that makes a generation reproducible. Same model, prompt and seed → same image.

CFG / guidance. How strictly the model follows the prompt. Too low: ignores you. Too high: burnt, over-saturated images.

Inpainting. Regenerating a selected region of an image — fixing a hand, changing clothing — while keeping the rest.

Upscaling / enhancement. Increasing resolution and detail after generation. The creativity setting decides whether the enhancer may reinterpret; too high redraws anatomy (Krea).

Reference conditioning. Generating with one or more reference images so identity, style or layout carries over. The best character-consistency tool in open models (Flux).

Character LoRA. A LoRA trained on one character’s images so every generation shows the same person. The heavier, stronger alternative to reference conditioning.

Video generation

T2V / I2V. Text-to-video and image-to-video. For adult work i2v is the workflow: the model animates a reference frame instead of inventing a subject — see the reference frame guide.

Reference frame. The still image an i2v model animates. Its sharpness, lighting and pose decide more of the result than the prompt.

Multishot. One generation producing several connected shots with the same character across cuts (Seedance 2.5, LTX-2.5).

Video extension. Generating a continuation from a clip’s last frame to lengthen it. How 5-second models make 15-second scenes.

Prompt expansion. The host rewriting your short prompt into detailed model-friendly language before generation. Good hosts tune this per use case.

Identity drift. The subject’s face or body slowly changing across a clip — the standard failure past ~8 seconds on most models (why AI porn looks off).

World model. Marketing-adjacent term for video models trained to simulate consistent scenes and physics rather than just frames; how Lightricks describes LTX.

Chat

System prompt. The hidden instruction that defines who the AI is: persona, register, boundaries. In companion products, this is where the character lives.

Persona / character card. The description a platform (or you) writes for a character; the system prompt is built from it.

Context window. How much conversation the model can see at once. Beyond it, things are forgotten unless the platform keeps summarised memory.

Roleplay tune vs writer tune. Roleplay tunes (Gemma 4) are trained for short, reactive, in-character chat; writer tunes (Qwen 3.5) for long consistent scenes. Good platforms route between them.

RLHF. Training a model against human preference ratings. It is how assistants got helpful — and how refusals got baked in deep enough that “jailbreaks” die (GPT Image 2).

Refusal. The trained “I can’t help with that.” Removed by fine-tuning or abliteration in open models; unremovable in closed ones.

Filters, policy and law

Vendor filter / policy layer. The moderation a host wraps around a model: prompt classifier before generation, image classifier on uploads, output scan after. The model underneath is often capable; the layer decides — the theme of the Seedance and Kling analyses.

Prompt classifier. A language model that reads your prompt’s intent. The reason euphemisms and misspellings stopped working everywhere in 2025.

Uncensored host. A platform that runs a capable model without a blocking policy layer for legal adult content — while still enforcing hard limits (minors, real-person NCII) at every level.

NCII. Non-consensual intimate imagery — real people, without consent, AI-generated included. A federal crime in the US under the Take It Down Act and illegal in most jurisdictions; see AI porn and the law.

Deepfake. AI content depicting a real, identifiable person. In the EU, labelling becomes mandatory under the AI Act’s transparency rules from August 2026.

Age verification / age assurance. Proving users are adults, increasingly required by law (UK Online Safety Act enforcement began 2025). Facial estimation, ID matching and similar methods.

Watermarking / machine-readable marking. Embedding a detectable signal that content is AI-generated; the EU’s Code of Practice standardises it.

Related in the Lab

• Stable Diffusion for NSFW: SDXL vs Pony vs Illustrious

• Best AI video models for NSFW in 2026

• AI porn and the law in 2026

Key takeaways

• Most confusion in adult AI comes from four pairs: open weights vs uncensored, checkpoint vs LoRA, fine-tuned vs abliterated, model vs vendor filter.

• If you internalise one idea, make it this: capability lives in the model, permission lives in the layer around it.

• Every definition above links to the article that goes deeper.

Living document — updated as the vocabulary moves. Missing a term? Tell us on X.