Prompt Counter

The token figure is an estimate, and deliberately a range. A real count needs the tokenizer of the model that will read the prompt, and those disagree with each other — an image model's text encoder is not GPT's. A single number labelled “tokens” would be precise for no model anyone is using, so this brackets it instead. One estimate counts words plus some punctuation, the other uses the four-characters-per-token rule of thumb, and whichever comes out smaller is the low end — on very short words the two swap places.

The verdict is read against 75 tokens, not 77. CLIP's context is 77 positions and two of them are never yours — the tokenizer puts a start and an end marker around whatever you write — so 75 is the budget a prompt actually has.

The three bands are encoder families, not a model directory. A prompt that is over on CLIP is usually fine on T5 and unconstrained on a multimodal model, which is why one verdict was never enough — and why there is no price column here: what a generation costs is a live figure on Visxy's own pricing, never a number a tool page made up.

No spaces is what a character-limited field actually counts; clauses are the comma-separated fragments most image prompts are written as; and words are whitespace-separated, which describes English rather than every script.

Every figure above is its own copy button, and Download saves the lot as a text file with the prompt attached — a count without the text it counted is unusable a week later.