How to Phrase a Request So the Model Gives You the Frame You Want
A request to a generative model is neither an incantation nor a search query. It is a description of a frame, and it works better the closer its structure is to how a photographer would describe one.
Order matters
In most models, words near the beginning carry more weight. A sensible order: who and what first, then pose and angle, then light, then surroundings, and technical characteristics last.
Specifics instead of judgements
«Beautiful» and «professional» carry no information: the model does not know your idea of beauty. «Soft side light from a window on the left» does. Replace evaluative words with descriptive ones and the result becomes predictable.
Do not describe absence
The phrase «without glasses» often produces glasses: the model sees the word «glasses» and responds to it. Exclusion has its own mechanism — a negative description — and that is what should be used, not a negation inside the main request.
Length
A long request is not a precise one. Past a certain volume the model's attention smears and the last elements of the description are simply ignored. A short exact description plus several iterations usually beats one two-hundred-word paragraph.
A working method
- Fix the generation seed
- Change one element of the description at a time
- Watch what actually changed
- Within a few steps it becomes clear which words the model responds to and which it skips
A limit you cannot get around
If a concept was absent from the training data, no phrasing will summon it. Having hit that wall, it is easier to change approach — paint the missing part separately, or photograph it — than to hunt for a magic word.
Go to the site →Read next
Local Models or Cloud Services: Which to Choose LoRA: Training a Model on a Particular Likeness Nude Photography: The Genre, Its History and Its Boundaries Light in Nude Photography: How It Sculpts FormThis material is intended for adults over 18 and is informational in nature.