Overview
This glossary defines the foundational mechanics of generative visual AI. Understanding these terms is essential for controlling your image generation workflow.
1. Latent Space
Definition: A compressed, multi-dimensional mathematical environment where the AI model stores and processes data representations.
The Mechanics: Unlike standard image editing, which manipulates pixels (red, green, blue values), generative AI operates in "latent space." The model compresses an image's features—shapes, colors, textures—into a dense vector representation. When you prompt a model, you are navigating coordinates within this space to reconstruct a new image from those vectors.
Why it Matters: Understanding latent space explains why AI can "dream" up combinations that don't exist (e.g., "a cyberpunk toaster"). It isn't pasting a toaster onto a background; it is mathematically blending the concept of "cyberpunk" and "toaster" in the abstract feature space before decoding it back into pixels.
2. CFG Scale (Classifier-Free Guidance)
Definition: A parameter that controls how strictly the generation model adheres to your text prompt versus its own internal probability distribution.
The Mechanics:
- Low CFG (1–6): The model is given more "creative freedom." It prioritizes coherent image structure over exact prompt matching. Useful for artistic or abstract styles.
- High CFG (7–15+): The model is forced to follow the prompt rigidly. This often results in higher accuracy for specific details but can introduce visual artifacts or "burn" (oversaturation/distortion) if pushed too high.
Best Practice: Start at CFG 7.0 for a balanced baseline. Increase for prompt fidelity; decrease for artistic variety.
3. ControlNet
Definition: A neural network structure that adds spatial conditioning to diffusion models, allowing for precise control over composition, pose, and structure.
The Mechanics: Standard diffusion models rely on text prompts, which are poor at describing exact spatial arrangements (e.g., "person standing on the left leg, arm raised 45 degrees"). ControlNet locks the structure of an input image (using edge detection, depth maps, or pose estimation) and forces the AI to generate the new image onto that specific "skeleton."
Use Cases:
- Replicating human poses from a reference photo.
- Keeping the exact layout of a room while changing the interior design style.
- Generating logos that strictly follow a black-and-white sketch.
4. LoRA (Low-Rank Adaptation)
Definition: A modular training technique that fine-tunes a large model on a specific concept (style, character, or object) without retraining the entire network.
The Mechanics: Training a full model (like SDXL or Flux) requires massive compute. LoRA injects small, trainable rank decomposition matrices into the model's existing layers. These "adapter" files are tiny (megabytes vs. gigabytes) but powerful enough to shift the model's output toward a specific aesthetic or subject.
Why it Matters: LoRAs allow for modular workflows. You can load a "base" model and attach a "Watercolor LoRA" and a "Robotics LoRA" simultaneously to blend those specific concepts efficiently.