The 77-token limit: what happens to a long prompt

SD 1.5 and SDXL encode text with CLIP, which reads at most 77 tokens at a time (75 usable, plus start and end tokens). A token is roughly a short word or a piece of a long one; punctuation and weights count too. What the tools do past 75 - A1111 splits the prompt into 75-token chunks, encodes each, and concatenates them. It shows a counter (for example 74/75) in the corner of the prompt box. - ComfyUI's text encode node also concatenates chunks. - Some other front ends simply truncate. Anything past the limit is silently ignored. Why it matters Each chunk is encoded separately, so words in chunk two do not "see" chunk one. "A woman in a red dress" split across a chunk boundary can become a woman and, somewhere, red. The BREAK keyword (A1111) Typing BREAK forces a new chunk at that point. Use it on purpose: subject and clothing in chunk one, BREAK, background and lighting in chunk two. That keeps related words together. Method 1. Write the prompt. 2. Watch the counter. If it is over 75, decide where the boundary falls, and put BREAK at a sensible place. 3. Cut tags that duplicate each other: "beautiful, gorgeous, stunning" spends three tokens on one idea. The practical answer If your prompt is 200 tokens, the model is not reading it the way you wrote it. Shorter and ordered nearly always wins.