AI Terms Glossary for Sellers: Key Words Explained Simply

The moment you open an AI tool to make a photo for a product listing, you get hit with prompts, seeds, tokens, and Elo scores all at once. I pulled together an AI terms glossary covering what sellers run into most, explained in plain English, no jargon required. This isn't for a quiz. It's so you actually know what's going on when a guide or a settings panel throws terminology at you.

Talking to the model

Prompt

A prompt is the text you type to tell the model what you want. The more specific it is, the closer your result lands to what you pictured.

Negative prompt

A negative prompt lists what shouldn't show up: extra fingers, garbled text, a logo that isn't yours. Not every model supports this, but when it does, it cuts down on bad takes.

Seed

A seed is the number a model starts from when generating. Same prompt, same seed, same result every time; change the seed and you get a new version of that same scene. This is handy once you've landed on a shot you like and just want to tweak small details.

Inference

Inference is the actual generation step, when the model turns your prompt into an image or video. When someone mentions "8 steps," they're talking about inference steps.

Money and limits

Tokens and credits

Tokens and credits are the internal currency these tools run on. You spend them on every generation, and different models drain your balance at different speeds. I broke down exactly what gets charged and how to stop wasting credits in a separate piece on credits and tokens in AI tools.

Resolution and 4K

Resolution is the pixel size of your image or video. 4K looks sharp, but it costs more and renders slower; for a product listing, something smaller is often all you need.

Upscaling

Upscaling takes a finished image or video and bumps it to a higher resolution as a separate step. It's cheaper than rendering straight to 4K, though the quality of the upscale itself varies quite a bit.

The models themselves

Checkpoint

A checkpoint is a specific version of a model, a snapshot frozen after training. Every time a new version ships, that's a new checkpoint.

Fine-tune

Fine-tuning retrains a model for a narrow job or style, say, matching a brand's visual identity. This mostly works on open-weight models.

LoRA

A LoRA is a lightweight add-on that layers a particular style or subject onto a model without retraining the whole thing. It's a small file that changes how a much larger model behaves.

Open weights

An open-weight model has its files published for anyone to download, run on their own hardware, and fine-tune. That's the opposite of closed models, which only run through an API. I went deeper on this in a piece on open-weight image models.

How people actually measure quality

Elo and leaderboards

Elo scoring compares models blind: people see two generations from the same prompt, pick the one they like better, and the wins get tallied into an Elo rating, the same math chess rankings use. Honestly, I wouldn't put too much weight on the number by itself. It reflects average scenes, not your specific shot. For a walkthrough on how to read these tables, I wrote about an AI image leaderboard.

Benchmark

A benchmark is a comparative test run across a set of tasks. It gives you a rough sense of a model's overall strength, but it won't tell you how your particular image will turn out.

Video and sound terms

Text-to-video and image-to-video

Text-to-video builds a clip from a written prompt. Image-to-video brings a still photo to life. For a product listing, you'll usually want the second mode, since it keeps your actual item intact instead of generating a new one.

Lip sync

Lip sync matches a character's mouth movements to spoken audio, though it still misses the mark on longer or fast-paced lines.

Native audio

Native audio means the model generates sound alongside the video itself, rather than having it layered on afterward.

Where it came from

Watermark and metadata

An invisible watermark embedded in the pixels (SynthID is one example) and metadata describing a file's origin (C2PA) both let someone check whether a piece of content was AI-generated. I covered this in more detail in a piece on labeling AI content.

All of this clicks a lot faster once you're actually doing it than reading a definition. Open a model on Flami, change the prompt, the seed, the resolution, and watch what shifts.

Frequently asked questions

What do prompt and seed mean in plain terms?

A prompt is the text description of what you want the model to make. A seed is the starting number for a generation: the same prompt plus the same seed gives you a repeatable image, while changing the seed produces a new variation of that scene.

What's the difference between open and closed models?

Open-weight models have their files published for anyone to download, run locally, and fine-tune. Closed models only run through the developer's API, so there's no way to retrain them for your own style.

What are tokens and credits in AI tools?

They're the internal currency these services run on, charged per generation. Different models and resolutions cost different amounts, so your balance doesn't drain at a steady rate. I've got a full breakdown of exactly what you're paying for in a separate article.

Why should a seller bother learning AI terminology?

So guides and settings panels actually make sense, and you're not burning credits on regenerations you didn't need. Once you know what a negative prompt, a seed, or upscaling actually does, you tend to land the right shot on the first or second try instead of guessing.

Sources

1. Flami: credits and tokens in AI tools
2. Flami: open-weight image models