← sources

Any tricks to handle paragraph-long image prompts without attention drop-off?

source: reddit (reddit:PromptEngineering)published 2026-11-14 12:00ingested 2026-11-15 02:36

retained text

It is annoying when you spend so much time writing extremely detailed prompts for your images, and then it just won’t work, it just looks like the model doesn’t even understand your instructions.   
  
I tried to be as detailed as possible, describing backgrounds in fine details, giving specific lighting instructions, subject placement all that good stuff. But half way through, the model kinda gives up. They read the first 80 tokens, find a few keywords it understands, and the rest just gets properly ignored.  
  
Looking for workarounds and saw the specs for that new MiniMax H3-Image model. It is open source with a token limit of 4096, sounds impressive. But does a model actually understand 4000 tokens tho? That seems a lot.  
  
Doc says it's built for precise instruction following and for understanding complex, detailed requirements. But wondering if anyone tested if it actually pays attention to the whole prompt that long?

document fetched 2026-11-15 00:34 · canonical https://reddit.com/r/PromptEngineering/comments/1xw0qq8/any_tricks_to_handle_paragraphlong_image_prompts