This guide explains advanced prompting techniques for better control over image generation on SeaArt.AI, applicable to all models except Flux, which uses a different system.
Last updated
You have learned how to write basic prompt with weights on Complete Prompting Guide. This guide will teach you advanced prompting techniques to have more control over image generation. Effects can show difference for each model but these rules are applies to all models work on . Only Flux doesn't work with these tools because it has different prompt understanding system.
Term keyword can be any word in your prompt. Keyword can be more than one word in some syntax.
Term factor is value of sampling steps. Scales between 1 and highest sampling step.
Term factor% is percentage value of sampling steps. Can't be higher that 1 (100%) and scales between 0-1.
Generation starts with keyword1 and changes to keyword2 after factor step.
Lower values sometimes will not have visible effect because it has small part of generation.
Order matters, first keyword1 is your starting term while keyword2 is secondary part of generation.
portrait of a girl, [smiling:frown:0.7] face with 20 step generation will be generated as:
Step 1-14 portrait of a girl, smiling face
Step 15-20 portrait of a girl, frown face
Images were generated with same settings with 20 steps. factor% changed with 0.1, 0.3, 0.5, 0.7, 0.9
As you can see first image has more angry face because only 2 step generated with smiling while last image has smiling face because 18 step generated with smiling and changed into frown at last 2 steps.
Generation starts without keyword and adds keyword to prompt after factor step.
Higher values will not have visible effect because main composition of image will be generated already and AI cannot make this type significant changes.
portrait of a girl, green eyes, [pink hair:7] with 20 step generation will be generated as:
Step 1-7 portrait of a girl, green eyes,
Step 8-20 portrait of a girl, green eyes, pink hair
Images were generated with same settings with 20 steps. Term factor is changed with 3, 7, 10, 13, 17
As you can see, there is no pinkish hair in the third and later images because AI already generated main composition and couldn't make changes even we add pink hair later. White color was totally luck, it could be any color includes pink.
Generation starts with keyword and removes keyword from prompt after factor step.
Higher values will not have significant differences because main composition will be generated with keyword already.
portrait of a girl, [ginger::3], black dress with 20 step generation will be generated as:
Step 1-3 portrait of a girl, ginger, black dress
Step 4-20 portrait of a girl, , black dress
Images were generated with same settings with 20 steps. Term factor is changed with 3, 5, 7, 10, 13
As you can see there is no significant difference after third image because main composition was already generated when we remove keyword from prompt. Blonde hair was totally luck, it could be any hair color includes ginger.
This syntax will swap keywords on each step. keyword can be more than two, you can add keyword as much as sampling steps but it's not recommended.
Lower keyword count equals to better results.
portrait of a girl, [blonde|ginger|black] hair with 20 step generation will be generated as:
Step 1-4-7-10-13-16-19 portrait of a girl, [blonde|ginger|black] hair
Step 2-5-8-11-14-17-20 portrait of a girl, [blonde|ginger|black] hair
Step 3-6-9-12-15-18 portrait of a girl, [blonde|ginger|black] hair
portrait of a girl, [blonde|ginger|black] hair
[frog|bird|cat]
Slash can be used to use actual parantheses in prompt without weight. Some tags include parantheses () and this syntax help you to use them.
mona (genshin impact) will be tokenized as a mona with 1x weight and genshin impact with 1.1x weight.
mona \(genshin impact\) will be tokenized as itself, without weighting and with actual parantheses.
(masterpiece, best quality), mona \(genshin impact\)
