The token limit that eats long prompts

Seventy-five words that count

The CLIP text encoders used by the Stable Diffusion family accept exactly 77 token positions. Two are taken by start and end markers, leaving 75 for you. A token is roughly three-quarters of an English word, and less for unusual words, which get split into pieces.

So a prompt of about 55 to 60 ordinary English words fills the budget. Everything after that is, in the plain implementation, cut off and thrown away.

Nothing in the interface tells you. The generation succeeds. The image looks like your prompt, because the beginning of your prompt is the part that was read. People conclude the model ignored their later clauses out of preference. It did not receive them.

What tools do instead, which is not better

Most modern interfaces avoid the hard cut by chunking: they split the prompt into 75-token blocks, encode each separately, and concatenate or average the results. This is why you can paste a 300-word prompt into a WebUI and see it react to the end of it.

It is a workaround, not a fix, and it has a distinct signature. The chunks are encoded independently, so a clause split across a boundary loses its relation to the rest. a woman holding a in chunk one and red umbrella in chunk two do not compose. The tool usually shows the chunk count somewhere small — 75/150 means two chunks. When a long prompt behaves as if it half-understood you, look at that number.

Newer pipelines built on T5 or similar text models allow 256 or 512 tokens, and read them as one sequence. That is the real fix, and it is why long, careful, sentence-structured prompts became worth writing only recently.

How to find your limit in one minute

Append a token that would be unmistakable if read:

<your long prompt> , and a bright yellow rubber duck

Fix the seed. Generate. No duck means the tail is not reaching the model. Move the duck to the front and generate again to confirm that the model does render ducks. This test takes two generations and settles an argument people have for months.

Writing to a budget

Once you know you have 75 tokens, prompt writing becomes an editing problem, and editing is a skill you can actually improve at.

Cut, in this order:

1. Quality incantations. masterpiece, best quality, 8k, highly detailed, award winning is often 12 tokens doing almost nothing.
2. Redundant synonyms. beautiful, gorgeous, stunning are one concept in three costs.
3. Articles and connectives, but only on bag-of-concepts encoders. On a syntax-reading model these are load-bearing and cutting them is counterproductive.
4. Anything you could achieve with a setting. Aspect ratio, resolution and colour grade belong in parameters or post-processing, not in tokens.

What survives should be: the subject, its important attributes, the setting, the light, and the register. That fits comfortably.

The related limit on the other side

There is a mirror-image problem worth naming. Prompts that are too short leave the model to fill in everything, and what it fills in is the training-set average — the most typical face, the most typical lighting, the most typical composition. A three-word prompt does not give you a neutral image; it gives you the centre of the distribution, which is why every "photo of a woman" from an unmodified model looks like the same woman.

Between the two limits there is a band, roughly 20 to 60 tokens on a CLIP model and rather wider on a T5 one, where the prompt is specific enough to escape the average and short enough to be read. Almost all good prompts live there. It is worth counting your tokens once, on the model you use most, so you know where the edges are rather than guessing.

One consequence of the limit is worth stating for anyone working outside English. Tokenisers are trained on text corpora that are heavily English, so an English word is often a single token while the same word in Hindi, Bengali, Tamil or Arabic is split into five or six pieces. A prompt in Devanagari can exhaust a 75-token budget in fifteen words. This is a real disadvantage and it is not the writer's fault. The workable route today is to prompt in English for the scene and use references or a fine-tune for whatever the English caption vocabulary does not cover, which is an unsatisfying answer and an honest one. It will improve as models are trained on more multilingual captions, and it has improved already; it is not yet solved.