READ4 min
WORDS984
SECTIONS3
TYPEEXPLAINER
CHECKED25 AUG 26
TL;DR — THE SHORT VERSION
Plain definitions of the prompting terms used with image generators, video models and LLMs, with the page’s settings ranges for things like photorealism offered as working advice rather than vendor figures.
- Seeds make results repeatable. Reusing a prompt with the same seed lets you reproduce an output you liked, or iterate from it.
- Negative prompts are not universal. Stable Diffusion and ComfyUI pipelines have the field; FLUX.2 lacks it.
- Identity lock is prompt text. Not a setting: tell the video model to keep the same face in every frame.
- Temperature sets randomness. Lower is steadier, higher is more varied, but check the vendor first: on Gemini 3.x Google says to leave it at the default.
- Chain-of-thought matters less now. OpenAI calls it unnecessary on reasoning models, which do better with high-level guidance.
◈ A SECOND GLOSSARY EXISTS
This one covers prompting terms. The AI glossary covers the wider vocabulary — sixty terms, honestly defined.
Every technical term you'll encounter when working with AI image generators, video models, and LLMs — explained in plain language. Bookmark this page and come back whenever a term doesn't make sense.
Image Generation Terms
CFG Scale (Classifier-Free Guidance)
What it is: A setting that controls how closely the model follows your prompt versus how much creative freedom it takes. Low CFG (3–5): the model interprets loosely, more natural results. High CFG (8–12): the model follows your prompt rigidly, but can introduce artifacts and over-saturation. Most models work best between 5–7 for photorealism.
CFG SCALE · LOOSE TO RIGID
How closely the model follows your prompt, using the ranges in the entry above.
LOW · 3–5Loose interpretation, more natural results.
5–7Where most models work best for photorealism.
HIGH · 8–12Follows the prompt rigidly; can introduce artifacts and over-saturation.
Steps
What it is: The number of denoising iterations the model performs during generation. More steps generally means more refined output, but with diminishing returns above 25–30. For most workflows: 20 steps for drafts, 25–30 for final output. Going above 50 rarely improves quality and significantly increases generation time.
Seed
What it is: A number that controls the randomness of generation. The same prompt and seed, on the same model with the same settings, usually reproduces the same output, though some tools and hardware still introduce small differences. Different seeds produce different outputs from the same prompt. Save seeds of outputs you like — they let you reproduce or iterate from a known starting point.
Sampler
What it is: The algorithm the model uses to generate the image step by step. Different samplers produce different results even with identical settings. DPM++ 2M Karras, Euler a and DDIM are common choices in Stable Diffusion tools; which looks best depends on the model and the image, so compare them on your own prompt. An earlier version of this entry called DPM++ 2M Karras the best sampler for photorealism, with no source.
Negative Prompt
What it is: A list of things you don't want the model to include in the output. Works by steering generation away from those concepts. Support varies by model and it is not universal — Stable Diffusion and ComfyUI pipelines expose a negative-prompt field, while Black Forest Labs states plainly that FLUX.2 does not support negative prompts and advises describing what you want instead. Check your model before writing one.Black Forest Labs, FLUX.2 prompting guide, re-read at source 11 Sep 2026: "FLUX.2 does not support negative prompts." An earlier version of this glossary listed FLUX among the models where negative prompts are essential, contradicting this site’s own correction of 25 Aug 2026. Corrected 31 Aug 2026. Common entries: CGI, plastic skin, airbrushed, watermark, blurry, extra fingers, deformed anatomy.
Denoising Strength
What it is: Used in img2img workflows. Controls how much the model changes the input image. 0.0 = no change at all. 1.0 = ignore the input, generate fresh. For upscaling and refinement: 0.35–0.50 preserves composition while improving detail. For moderate changes such as lighting or style: 0.55–0.70. For major changes that keep the basic composition: 0.75–0.85. These ranges are this site’s working advice; no vendor publishes them.
img2img
What it is: Using an existing image as the starting point for a new generation, combined with a text prompt. The model modifies the input image according to the prompt at the denoising strength you set. Used for upscaling, style transfer, lighting changes, and refinement of AI-generated images.
Inpainting
What it is: Editing a specific region of an image while leaving the rest unchanged. You mask the area you want to change, write a prompt for what should replace it, and the model regenerates only that region. Used for removing watermarks, changing backgrounds, fixing specific details, or replacing elements in an existing image.
LoRA (Low-Rank Adaptation)
What it is: A small add-on model trained to reproduce a specific style, person, or subject that the base model doesn't know about. LoRAs are layered on top of the base model and are activated with a trigger word in the prompt. Used to lock in a specific person's face, apply a distinctive art style, or reproduce product-specific aesthetics.
ControlNet
What it is: A conditioning system that uses a reference image to control composition, pose, or depth in the generated output. Types include: OpenPose (matches a specific body pose), Depth (matches spatial depth from a reference), Canny (matches edges and lines). IP-Adapter is often used alongside ControlNet but is a separate adapter: it lets a reference image act as a prompt, carrying style or a face into the output.Ye et al., Tencent AI Lab, IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models, arXiv, Aug 2023, read at source 17 Sep 2026: “In this paper, we present IP-Adapter, an effective and lightweight adapter to achieve image prompt capability for the pretrained text-to-image diffusion models.” An earlier version of this entry listed IP-Adapter as a ControlNet type.
TAKEAWAY
Treat the settings ranges here as starting points, not rules, and check what your own model supports before relying on a field such as a negative prompt.
Video Generation Terms
Temporal Consistency
What it is: How consistent the subject looks across frames in an AI-generated video. Poor temporal consistency means the character's face, clothing, or appearance changes between frames — a common failure mode in AI video. Improved by adding identity lock instructions: "same consistent face and features throughout every frame, no morphing."
Identity Lock
What it is: A technique where you explicitly instruct the video model to maintain a consistent subject identity across all frames. Not a model setting — it's a prompt instruction. Add it explicitly: "same face, no identity drift, same clothing and features throughout every frame, no morphing."
Bracket Format
What it is: A structured prompt format used for video generation that organizes instructions into labeled sections: [MOTION], [SUBJECT], [ENV], [CAMERA], [STYLE], [DURATION]. This site tested it in June 2026, before Seedance 2.5 was released. It is not the structure ByteDance publishes, and other vendors order prompts differently, so use it as a checklist and reorder for your model; see the anatomy of an AI video prompt. Keeping motion, subject and environment apart makes it easy to see that none is missing. An earlier version of this entry said the format was developed for Seedance 2.5 and works across most video models.
TAKEAWAY
Consistency in AI video comes from what you write: put the identity lock in the prompt itself, and keep motion, subject and setting in separate parts of it.
LLM Terms
System Prompt
What it is: Instructions given to an LLM before the conversation starts that define its role, behavior, constraints, and output format. Applied at the session level — the model refers back to it throughout the conversation. More powerful than a user prompt for controlling consistent behavior. Used to set personas, define tone, restrict topics, and specify output formats.
Temperature
What it is: A setting that controls how random the model’s word choices are. Lower values give steadier, less open-ended answers; higher values give more varied ones. Check the vendor before changing it. For Gemini 3.x models Google now strongly recommends leaving it at the default, and on 21 July 2026 it deprecated the setting in its API.Google, Prompt design strategies, read in a browser 22 Sep 2026: “Lower temperatures are good for prompts that require a more deterministic or less open-ended response, while higher temperatures can lead to more diverse or creative results.” And: “Although you can modify these parameters, we strongly recommend keeping them at their default values for Gemini 3.x models. Changing these parameters (for example, setting the temperature below 1.0) can cause unexpected behavior, such as looping or degraded performance, particularly in complex mathematical or reasoning tasks.” Corrected 22 Sep 2026: this entry said “Use low temperature for factual tasks, higher for creative writing” and called 0.7 a good default, with no source.
Context Window
What it is: The maximum amount of text an LLM can process in a single session — including your instructions, the conversation history, and any documents you've provided. Sizes differ by model and change often, so check the vendor’s model page rather than a glossary. When a conversation outgrows the window, what happens depends on the product: the request may fail, or older parts of the conversation may be cut or summarised to make room, and anything cut is no longer seen by the model. An earlier version of this entry gave a range of sizes and said the model “begins to forget”.
Chain-of-Thought Prompting
What it is: A technique that instructs the LLM to reason through a problem step by step before answering, via phrases like "think step by step." Largely superseded. It gave large accuracy gains on single-pass models, but reasoning models now plan internally before answering, so OpenAI calls the instruction unnecessary on them — it “may not enhance performance (and can sometimes hinder it)”. OpenAI says reasoning models do better “on tasks with only high-level guidance”.OpenAI, Reasoning best practices and Prompt engineering, read at source 16 Sep 2026. An earlier version of this entry said the instruction can encourage "over-deliberation" and is still useful on fast models; neither is in OpenAI's guides.
TAKEAWAY
Set lasting behaviour in the system prompt, check each vendor’s advice before touching temperature, and on reasoning models give high-level guidance rather than asking them to think step by step.
More prompts, when something changes.
Prompts, model guides and workflow notes, sent when there is something new. Free.
SUBSCRIBE FREE ↗