AI Image Generation Glossary: Complete Terminology Guide
AI image generation terminology encompasses the technical vocabulary, concepts, and jargon used in artificial intelligence image creation. This glossary provides clear definitions for beginners and professionals working with GLMImage, Stable Diffusion, DALL-E, and other AI image generators.
Key Takeaways
-
Understanding core terminology makes you a better prompter—knowing what "guidance scale," "CFG," and "steps" actually do is the difference between tweaking settings and hoping.
-
AI image generation borrows terminology from multiple fields: machine learning, photography, graphic design, and art history—creating a hybrid vocabulary that can confuse newcomers.
-
This glossary focuses on practical usage—definitions emphasize how each term affects image generation quality and workflow, not just theoretical meaning.
A
Aspect Ratio
The proportional relationship between an image's width and height, expressed as two numbers separated by a colon (e.g., 16:9, 1:1, 9:16). Determines image shape regardless of resolution.
Common Aspect Ratios:
- 1:1 - Square (Instagram posts)
- 16:9 - Horizontal (YouTube thumbnails)
- 9:16 - Vertical (TikTok, Stories)
- 4:3 - Standard photo
- 3:2 - Traditional camera
Autoregressive Model
A type of AI model that generates sequences by predicting the next element based on previous elements. GLM-Image uses a 9B autoregressive component for semantic understanding before visual generation.
Why It Matters: Autoregressive components enable text understanding and coherent image generation by "thinking through" the content before creating.
AI (Artificial Intelligence)
Computer systems capable of performing tasks that typically require human intelligence, including visual perception, language understanding, and creative generation.
B
Batch Size
The number of images generated simultaneously in a single operation. Higher batch sizes improve throughput but require more memory.
Trade-offs:
- Small batch (1-2): Less memory, slower overall
- Medium batch (4-8): Good balance
- Large batch (16+): Fastest, needs more VRAM
Bias (AI)
Systematic error in AI outputs that reflects training data prejudices. Can result in stereotypical or unrepresentative generated images.
Example: AI may generate certain professions predominantly as one gender due to training data bias.
Bleed
When visual elements from one part of an image unintentionally affect another area, creating artifacts or unwanted color shifts.
Blueprint
Detailed planning document or reference image used as a foundation for AI generation, especially in image-to-image workflows.
C
CFG (Classifier-Free Guidance)
A technique that uses both conditional and unconditional predictions to guide image generation toward the prompt while maintaining quality. The strength is controlled by guidance scale.
How It Works: The model generates both with and without prompt guidance, then combines them based on the guidance scale value.
Checkpoint
A saved state of a trained model at a specific point in training. Users can load different checkpoints to achieve different styles or capabilities.
Usage: "Load the realistic photography checkpoint for portrait generation."
CLIP (Contrastive Language-Image Pre-training)
OpenAI's model that connects text and images, widely used in AI image generation for understanding prompt-text relationships.
Compositing
Combining multiple visual elements from different images into a single cohesive image.
Concept Art
Illustrations that convey ideas for characters, environments, or objects before final production. AI accelerates concept art creation significantly.
Content Policy
Rules defining what content an AI model can and cannot generate, typically prohibiting violence, hate speech, sexual content, and copyrighted characters.
Cross-Attention
A mechanism that allows text prompts to influence specific parts of image generation, enabling precise control over visual elements.
D
Denoising
The process of removing noise from an image. Diffusion models generate images by starting with random noise and progressively denoising toward a coherent image.
Steps: More denoising steps = higher quality but slower generation.
Diffusion Model
A type of generative model that learns by adding noise to images and then learning to reverse the process. GLM-Image uses a 7B diffusion decoder.
How It Works: Model trains on noisy/clean image pairs, then generates by starting from random noise and progressively denoising.
Digital Art
Artwork created using digital tools, including AI-generated images. Can refer to specific aesthetic styles (e.g., "digital art style" prompt modifier).
Detail Enhancement
Techniques for increasing fine details in generated images, including prompt keywords ("highly detailed," "sharp focus") and upscaling methods.
DPM (Diffusion Probistic Models)
A family of advanced sampling methods for diffusion models that can improve quality or reduce steps needed.
DALL-E
OpenAI's series of text-to-image models (DALL-E 1, DALL-E 2, DALL-E 3). DALL-E 3 has strong English text rendering but poor Chinese character support.
E
Embedding
A numerical representation of data (text or images) that captures semantic meaning in a lower-dimensional space. Used for prompt understanding and style transfer.
Encoder
The component of an AI model that converts input data (text prompts or reference images) into internal representations for processing.
Epoch
One complete pass through a training dataset. Models are trained over multiple epochs to improve learning.
ESRGAN (Enhanced Super-Resolution GAN)
A popular upscaling model that can increase image resolution while enhancing details.
Expanding (Image)
Increasing canvas size while extending the image content outward, creating new content that matches the original style.
F
FID (Fréchet Inception Distance)
A metric for evaluating generated image quality by comparing distributions of real and generated images. Lower scores indicate better quality.
Scores: FID below 20 is generally considered good.
Fine-tuning
The process of further training a pre-trained model on specific data to adapt it for particular styles or subjects.
FLUX.1
Black Forest Labs' 12B parameter open-source image generation model with strong photorealism capabilities but weaker Chinese text rendering.
FPS (Frames Per Second)
Measurement of generation speed or video frame rate. In image generation, refers to images generated per second in batch processing.
Full Precision
Using 32-bit floating-point numbers for model calculations. Most accurate but requires maximum memory.
G
GAN (Generative Adversarial Network)
A type of generative model using two networks (generator and discriminator) that compete against each other. Less common in modern image generation than diffusion models.
Generation Time
The time required to create an image from a prompt, typically 8-30 seconds depending on model, hardware, and resolution.
GLM-Image
Zhipu AI's 16B parameter hybrid image generation model combining a 9B autoregressive component with a 7B diffusion decoder. Industry-leading for bilingual text rendering.
Guidance Scale
A parameter (typically 1.0-20.0) controlling how strongly the model follows the text prompt. Higher values = more prompt adherence, lower values = more creativity/randomness.
Recommended Range: 7.5-12.0 for most applications
GPU (Graphics Processing Unit)
Specialized hardware that accelerates AI image generation. Essential for local model inference.
Requirements: 12-24GB VRAM for most current models.
H
Hallucination
When an AI model generates content not present in or implied by the input, particularly problematic for text rendering.
Hybrid Architecture
A model combining multiple AI approaches. GLM-Image hybridizes autoregressive language modeling with diffusion generation.
High-Resolution Fix
A technique for generating images at lower resolution then upscaling and adding details, allowing higher quality without excessive memory usage.
I
Image-to-Image (Img2Img)
Using an existing image as input reference or starting point for generation, with a "strength" parameter controlling how much the original is preserved.
Strength Values:
- 0.1-0.3: Minor adjustments
- 0.4-0.5: Noticeable changes
- 0.6-0.8: Significant transformation
- 0.9+: Almost complete regeneration
Inpainting
Filling in missing or selected parts of an image with AI-generated content that matches the surrounding area.
Inference
The process of using a trained model to generate outputs (images) rather than training the model.
Iterative Generation
Generating multiple versions of an image, refining prompts and settings based on previous results.
K
kCHASE
A sampling method that restricts token generation to the k most likely next tokens, reducing randomness.
Key Frame
A reference image defining style or composition for generating variations or animation frames.
L
LCM (Latent Consistency Models)
A technique for reducing diffusion steps while maintaining quality, enabling faster generation.
LLM (Large Language Model)
A large-scale text model like GPT. GLM-Image incorporates LLM components for text understanding.
LORA (Low-Rank Adaptation)
A lightweight fine-tuning method that can adapt models to specific styles or subjects with minimal training. Popular in the Stable Diffusion ecosystem.
Latent Space
The compressed representation of images used internally by diffusion models during generation.
LoRA (Low-Rank Adaptation)
A parameter-efficient fine-tuning technique that allows adapting large models to specific styles or subjects with minimal additional training.
M
Midjourney
A commercial AI image generation service accessed through Discord. Known for high artistic quality but requires subscription and has poor text rendering.
Mixed Precision
Using both 16-bit and 32-bit floating-point numbers to reduce memory usage with minimal quality loss.
Model Card
Documentation describing a model's intended use, limitations, training data, and evaluation results.
Modelfile
Configuration file defining how a model should be run, including parameters, templates, and system prompts.
N
Negative Prompt
Text describing what to avoid or exclude from generation. Used to reduce unwanted elements.
Example: "blurry, distorted, extra limbs, watermark, text"
Noise
Random variation added to images during diffusion model training. Generation reverses this noise process.
Nucleus Sampling (top_p)
A sampling method that selects from the smallest set of tokens whose cumulative probability exceeds threshold p.
O
Open Source
Model with publicly available code and weights that can be used, modified, and distributed (often with license conditions).
Outpainting
Extending an image beyond its original borders, generating new content that matches the existing image style and composition.
Overfitting
When a model learns training data too specifically, failing to generalize. Can cause models to only reproduce training examples.
P
Parameters
The internal weights and values a model learns during training. More parameters generally enable more capability but require more memory and computation.
GLM-Image: 16B parameters (9B autoregressive + 7B diffusion)
Photorealistic
An image quality resembling a photograph rather than artwork or illustration.
Pipeline
The complete workflow from input to output, often implemented as reusable code.
Prompt
The text input describing what image to generate. Quality of prompts significantly affects output quality.
Prompt Engineering
The practice of crafting effective prompts to achieve desired outputs, including structure, keywords, and modifiers.
PRC (People's Republic of China)
Geographic context for Chinese language support, important for text rendering capabilities.
Q
Quantization
Reducing model precision (e.g., 32-bit to 4-bit) to decrease memory usage and increase speed, with minimal quality loss.
Common Levels: FP32, FP16, INT8, INT4 (4-bit)
R
Resolution
The dimensions of an image in pixels (e.g., 1024×1024). Higher resolution = more detail but slower generation.
RLHF (Reinforcement Learning from Human Feedback)
Training method using human preferences to align model outputs with user expectations.
RNN (Recurrent Neural Network)
An earlier neural network architecture for sequential data, largely superseded by transformers in modern models.
S
Sampler
The algorithm that determines how the model generates each step of the diffusion process. Different samplers have different speed/quality trade-offs.
SDXL (Stable Diffusion XL)
Stability AI's 6.6B parameter image generation model. Has the largest ecosystem but weaker text rendering than GLM-Image.
Seed
A numerical value that initializes the random number generation for reproducible results. Using the same seed with the same prompt produces identical outputs.
Steps
The number of denoising iterations in diffusion generation. More steps typically improve quality but take longer.
Recommended: 30-50 steps for most models
Stable Diffusion
A series of open-source image generation models from Stability AI, with a large community ecosystem of tools and modifications.
Style Transfer
Applying the artistic style of one image to the content of another, achievable through image-to-image workflows.
Style Modifier
Keywords added to prompts to influence visual style (e.g., "oil painting," "digital art," "photorealistic").
T
Temperature
A parameter controlling randomness in generation. Lower values = more deterministic, higher values = more random/creative.
Typical Range: 0.1-2.0, with 1.0 as default
Text Encoding
Converting text prompts into numerical representations the model can process.
Text Rendering
The model's ability to generate accurate, legible text within images. GLM-Image excels with 95% English and 90% Chinese accuracy.
Text-to-Image (Txt2Img)
Generating images from text descriptions without a reference image.
Token
The smallest unit of text that a model processes. Models work with tokens rather than individual characters.
Example: "apple" might be one token, while "international" might be two.
Transformer
A neural network architecture using attention mechanisms, foundational to modern language models and some image generation approaches.
U
Uncensored Model
A model trained without content safety filters, capable of generating content restricted in commercial models. Use varies by jurisdiction.
Upscaling
Increasing image resolution while maintaining or enhancing quality, often using AI models like ESRGAN.
Use Case
A specific application or scenario for AI image generation, such as "product photography" or "book cover design."
V
VAE (Variational Autoencoder)
A type of generative model that learns to compress and decompress data. Used in some diffusion models for latent space representation.
VRAM (Video RAM)
Memory on a GPU used for model inference and image generation. More VRAM enables larger or higher-quality models.
Minimums: 12GB for GLM-Image (optimized), 8GB for SDXL
Variation
Generating multiple images from the same or similar prompts to explore different interpretations.
W
Weights
The learned parameters of a neural network that determine how it processes inputs. Model files consist primarily of weight values.
Workflow
A repeatable process for achieving specific types of results, often combining multiple techniques or steps.
X
XTL (Extended Token Language)
A hypothetical or specialized token format for handling specific character sets or languages.
Y
YOLO (You Only Look Once)
A real-time object detection model, sometimes mentioned in AI image contexts but not directly related to generation.
Z
Zero-Shot
The ability of a model to perform tasks without specific training examples, generalizing from its broader training.
Zhipu AI
The Chinese AI company that developed GLM-Image and the GLM series of language models.
Acronym Quick Reference
| Acronym | Full Term | Relevance |
|---|---|---|
| AI | Artificial Intelligence | Core technology |
| AR | Autoregressive | Model type |
| CFG | Classifier-Free Guidance | Generation technique |
| FID | Fréchet Inception Distance | Quality metric |
| GAN | Generative Adversarial Network | Model architecture |
| GPU | Graphics Processing Unit | Hardware |
| LLM | Large Language Model | Text understanding |
| LoRA | Low-Rank Adaptation | Fine-tuning method |
| RLHF | Reinforcement Learning from Human Feedback | Training method |
| SDXL | Stable Diffusion XL | Competing model |
| VAE | Variational Autoencoder | Model component |
| VRAM | Video RAM | Hardware memory |
[See these terms in action — generate an image →](/general/image-generation/commercial-poster-generator)
[See these terms in action — generate an image →](/general/image-generation/commercial-poster-generator)
[See these terms in action — generate an image →](/general/image-generation/commercial-poster-generator)
Frequently Asked Questions
What does "CFG" mean in AI image generation?
Classifier-Free Guidance — the setting that controls how strictly the model follows your prompt. Low values give the model freedom; high values force literal adherence, often at the cost of image quality.
What is the difference between text-to-image and image-to-image?
Text-to-image starts from your description alone. Image-to-image starts from a reference picture plus a prompt, and a strength setting decides how far the result can move from the reference.
What is a LoRA?
A small fine-tune file that teaches a base model a specific style, face, or object. They are common in the Stable Diffusion ecosystem.
Do I need to know these terms to use GLMImage?
No — the web interface handles the technical settings for you. The vocabulary matters when you read model comparisons or use developer tools.
