📚 Authoritative Lexicon • Google E-E-A-T Verified

A to Z Generative AI Glossary

Every essential technical term in prompt engineering, diffusion physics, and LLM sampling defined with actionable pro tips.

Image Diffusion [C]

CFG Scale (Classifier-Free Guidance)

A numerical parameter that dictates how strictly the AI diffusion model adheres to your prompt text versus freely improvising artistic variations. Standard values range from 3.5 to 7.0.

PRO TIP: Higher CFG (>10) can cause over-saturation, contrast burn, and plastic skin.

Architecture [C]

Checkpoint Model

A complete, pre-trained neural network weights file (such as SDXL 1.0 or Flux.1 Dev) capable of generating standalone imagery without external adapters.

PRO TIP: Checkpoints define the foundational aesthetic, anatomy comprehension, and resolution ceiling.

Spatial Control [C]

ControlNet

A neural network structure that provides explicit spatial conditioning (depth maps, edge detection, pose skeletons) to guide diffusion composition with pixel precision.

PRO TIP: Use OpenPose ControlNet to transfer complex human poses into Midjourney/SD renders.

LLM Engineering [F]

Few-Shot Prompting

The prompt engineering technique of providing 2 to 5 concrete input-output demonstration examples inside the prompt to guide formatting and tone consistency.

PRO TIP: Few-shot examples drastically reduce hallucinations in JSON and code generation.

Image Editing [I]

Inpainting

The process of masking a specific region of an image and regenerating ONLY that portion with a new prompt while preserving the surrounding composition.

PRO TIP: In Midjourney, this is accessed via the "Vary (Region)" button.

Deep Learning [L]

Latent Space

A compressed mathematical multi-dimensional space where conceptual relationships between visual and semantic tokens are mapped and navigated by AI models.

PRO TIP: Diffusion models generate images by denoising random gaussian noise inside latent space.

Model Fine-Tuning [L]

LoRA (Low-Rank Adaptation)

A lightweight adapter file (typically 10MB to 200MB) trained on top of a foundational checkpoint model to teach it a specific art style, character face, or clothing item.

PRO TIP: LoRAs load instantly without requiring retraining the massive base model.

Constraint Tuning [N]

Negative Prompting

Instructions provided to diffusion models (via the --no flag in Midjourney) that penalize unwanted visual artifacts like extra limbs, blurriness, or watermarks.

PRO TIP: Avoid over-filling negative prompts with dozens of terms; it dilutes the primary subject focus.

Image Editing [O]

Outpainting (Canvas Expansion)

Extending the borders of an existing image outwards into new canvas territory, allowing panoramic reveals while matching lighting and style.

PRO TIP: In Midjourney, use the arrow Pan buttons (⬅️ ➡️ ⬆️ ⬇️) or Custom Zoom.

Inference Efficiency [Q]

Quantization (GGUF / EXL2)

Compressing model weights from 16-bit floating point down to 4-bit or 8-bit integers, enabling massive LLMs like DeepSeek-R1 to run locally on consumer VRAM.

PRO TIP: Q4_K_M quantization retains 98% of intelligence while cutting memory usage in half.

Reproducibility [S]

Seed (Deterministic Entropy)

A starting numerical integer (e.g. --seed 42819) used to initialize the pseudo-random noise field. Reusing the exact same seed and prompt produces near-identical outputs.

PRO TIP: Keep the seed constant when testing minor word changes to isolate parameter effects.

LLM Sampling [T]

Temperature & Top-P

Hyperparameters controlling token selection randomness. Lower temperatures (0.2) force deterministic logical answers, while higher values (0.8) boost creative divergence.

PRO TIP: For code audits, keep temperature at 0.1; for storytelling, set to 0.75.

LLM Prompting [Z]

Zero-Shot Prompting

Prompting an AI model to perform a task directly without providing any prior examples or demonstration pairs, relying purely on pre-trained understanding.

PRO TIP: Zero-shot works well for broad summaries but falters on strict JSON schema outputs.