05 / GLOSSARY
AIGC Glossary
29 TERMS
Every concept in one short paragraph. Organised into basics, image, audio/video, and engineering, with the common English term kept alongside for reference.
Basics10
- PromptPrompt
- The instruction or description you give a generative model; it directly steers the direction and quality of the output. Stating subject, style, and constraints clearly beats stacking length.
- TokenToken
- The smallest unit of text a model processes — roughly one Chinese character or half an English word. Billing, context length, and rate limits are all counted in tokens.
- Context WindowContext Window
- The maximum number of tokens a model can process in one pass — how much material it can 'see'. Content beyond the window needs chunking, summarising, or retrieval.
- InferenceInference
- Running a trained model once to produce a result. Every 'generation' is an inference call; usage cost and latency live here.
- Fine-tuningFine-tuning
- Continuing to train a pretrained model on your own data so outputs match a specific style or task. LoRA is the most common lightweight method.
- Retrieval-Augmented GenerationRAG
- Retrieve relevant documents from an external knowledge base first, then have the model answer from them. Used to suppress hallucination and connect private knowledge.
- AgentAgent
- An AI system that plans steps, calls tools, and completes multi-step tasks autonomously. Coding agents and persistent personal assistants are the common forms.
- MultimodalMultimodal
- One model that understands or generates text, images, audio, and video together. Current Gemini and GPT flagship models are multimodal.
- HallucinationHallucination
- When a model states fabricated facts with confidence. High-stakes use calls for retrieved sources and cross-checking — prompt wording alone does not fix it.
- Open WeightsOpen Weights
- Model parameter files published publicly, free to self-host and fine-tune. Open weights do not automatically mean a commercial-use licence — check each repository's terms.
Image Generation09
- Diffusion ModelDiffusion
- The generate-by-denoising paradigm behind Stable Diffusion, FLUX, and Midjourney's image engines.
- Image-to-ImageImage-to-Image
- Generating a new image from a source image plus a prompt, preserving composition and palette. The strength parameter sets how far it deviates.
- InpaintingInpainting
- Repainting only a selected region while keeping the rest of the pixels untouched — the standard way to fix hands, swap backgrounds, or clear clutter.
- OutpaintingOutpainting
- Extending the canvas outward and generating new content — for changing aspect ratio or completing a cropped frame.
- UpscalingUpscaling
- Enlarging an image while reconstructing detail instead of stretching pixels. The standard step before print or large-screen use.
- ControlNetControlNet
- Conditioning generation on pose skeletons, edge maps, or depth so composition and posture stay under your control — the go-to for 'same pose, new style'.
- LoRALoRA
- A small style or character adapter trained from a few dozen images, plugged on top of a base model; files are usually tens of MB.
- SeedSeed
- The random number that fixes the initial noise. Keep the seed and vary the prompt to compare phrasings on a stable composition.
- Style TransferStyle Transfer
- Generation that keeps the content structure and swaps the visual style — one reference image restyles a whole new piece.
Audio & Video05
- Text-to-SpeechTTS
- Synthesising natural speech from text. Emotion tags, pause control, and multi-speaker handling separate the current generation from the last.
- Voice CloningVoice Cloning
- Replicating a voice from seconds to minutes of sample audio. Commercial use requires the speaker's consent — an industry red line, not a suggestion.
- Stem SeparationStem Separation
- Splitting a finished track into vocals, drums, bass, and more. The foundation of covers, remixes, and practice backing tracks.
- Lip SyncLip Sync
- Matching a character's mouth movement to given speech — the first gate of realism in digital-human video and dubbing workflows.
- Audio ReactivityAudio Reactivity
- Driving visuals from a track's beat and frequency bands. The core capability of music-video generators, usually mapped stem by stem.
Engineering & Practice05
- Rate LimitRate Limit
- The cap on calls or tokens per time window. Free tiers are tighter; automation scripts must handle 429 retries with backoff.
- SandboxSandbox
- An isolated execution environment so an agent can run code and edit files without touching the real system — the safety base of coding agents.
- Model Context ProtocolMCP
- An open standard for connecting tools and data sources to models — the USB-C of AI applications: adapt once, use everywhere.
- BenchmarkBenchmark
- Scored evaluations on fixed question sets. Watch for data contamination and leaderboard gaming; your own real cases are the final word.
- WatermarkWatermark
- An invisible, machine-verifiable provenance mark embedded in generated content (such as SynthID) — declares AI provenance without changing how it looks.
Definitions favour everyday usage over academic completeness; if a wording is off, please flag it via the contact page. Read the inclusion methodology →