c/perchance · by Jack_Fosset@lemmy.world · 1y

How to prevent "spilling description" of one element into another

When I am describing a scene in the prompt of https://perchance.org/ai-text-to-image-generator for several elements lets say three persons, I want to describe their characteristics and give them some features for example three women, first women has blonde hair, glasses and neck tattoo, second woman is brunette wearing pink hat and smoking and third woman is redhead with long hair, eyebrow piercing and sticks her tongue out What AI does is to mix this description between all the characters, I tried to add thing like does not smoke, does no wear glasses ect. to description of each character but that didn't have much effect. Any idea how to "border" description for each element?

1 points · 14 comments · view on lemmy.world

14 Comments

dannooppo@lemmy.world · 3 pts · 1y (4 replies)

The keyword BREAK, parenthesis () colon : semi-colon ; and dot . go a long way. While parenthesis are used for emphasis and weight, you can also use them to bundle parts of prompts. Colon to start a list, semi-colon to terminate. Won't always work as expected, but who doesn't like to fiddle.

Use BREAK to signal a new 'layer' in your image, essentially defining a new image within the image.

E.g. using art style 'casual photo' or 'professional photo' with prompt professional portrait of a female in a colorful dress BREAK a psychedelic black and white background will yield exactly that.

Below prompt, with 'professional photo' worked for me, consistently (okay, fairly), also with more than 1 image in one run:

professional full body photo of three slender lithe females BREAK one girl is a Caucasian redhead, casual wear BREAK one girl is African in a red dress, barefoot BREAK one girl is Japanse with white hair and big boots BREAK a psychedelic black and white background

The same run, 3 images:

As you can see, you will still need to be very specific between BREAKs to prevent spilling, but I think you will get the gist...

enthusiasm_headquarters@lemmy.world · 1 pts · 1y

Wow, @danno, all this really should be in the Tips and general guidance of perchance itself.

Jack_Fosset@lemmy.world · 1 pts · 1y (2 replies)

Thanks I will try to play with it, are these BREAK, parenthesis () colon : semi-colon ; and dot . some universal keywords and characters or that is just a thing that seems to work with perchance?

dannooppo@lemmy.world · 2 pts · 1y (1 reply)

They are (seem to be) universal

I know BREAK is and works like advertised.
The parentheses can be used to add more weight to a (piece) of a prompt. Given that no parentheses is a weight of 1, then a single () is 1.1 * 1 weight, (()) is 1.1 * 1.1 * 1 weight and so on. Or they can be used like: (some really important piece of prompt):1.21, adding a certain weight directly. Don't overdo the decimal, something between 1.1 and 1.7 will usually do the trick.
The colon : is often used like a sum-up, e.g. the female is wearing: ..., ..., ... and ... ending with a dot or a semi-colon.
The semi-colon ; is used to split part of a scene like they are dancing; optimal lighting, low camera angle; leisure time concept.

It's not an exact sciene, but it will help.

Jack_Fosset@lemmy.world · 1 pts · 1y

Interesting is that this works quite well in perchance but if I try to put what you wrote in prompt in comfyUI with standard text-to-image workflow regardless of a base mode used (SD 1.5, SD 3, SDXL 1, Pony, Ilustrious) result is an absolute mess, it generates on average from 2 to 5 characters, dresses, prompts and races are utterly random, I wonder what magic is used in perchance that results are so much accurate to prompt

dannooppo@lemmy.world · 2 pts · 1y (1 reply)

@Jack_Fosset@lemmy.world @Koto@lemmy.world @enthusiasm_headquarters@lemmy.world

This is a bit hard to find in the Perchance Hub and by now it's SD tips are a bit obsolete, but still enlightening. Search on the page for the 'BREAK keyword'.

This reddit post might come in handy too: Let's make some realistic humans: Now with SDXL [Tutorial] (again SD, but much is still relevant).

Jack_Fosset@lemmy.world · 1 pts · 1y

By now I looked into many tutorials and reddit posts and civitas pictures for inspiration but they mosty generate single character or two at best, mostly without carrying which prompt each character gets, and even of there is some picture with multiple characters someone post as example, when I read the prompt there is no proper description of them, so it was just a random seed generation they got, In general I start to think that Differentiate prompts for different characters is something the AI is not good at, and it is not the problem with how I write the prompt but in general how it works and how it learned, and thing like interactions between the characters and their positional relationship is something that AI from the prompt almost never understands and just mix it even worse than prompts.

Koto@lemmy.world · 1 pts · 1y (1 reply)

The AI's have come a long way, especially the recent versions that catch a lot of details, but they are just not there yet for the task you're describing working on the prompt alone. If you really want a detailed scene of three or more people you will need to dive into the world of LoRas, Control Net and other tools, in ComfyUI or similar program. Another approach would be to make a very detailed image of each of the three characters here on perchance, and then, ask Flux.1 Kontext to put them all in one scene doing something you want without changing their appearance. I've had very good results with this approach, and it's much easier, too.

Jack_Fosset@lemmy.world · 1 pts · 1y

Thank, I never heard of Flux.1, I will have a look, but that's for time when I learn more with comfyUI in general, so far I liked the simplicity of perchance, but I think I am starting to hit its limitations

yotv@lemmy.world · 1 pts · 1y

i am not sure giving the Character names can work sometimes but not always .i noticed it being worse today. maybe its just a coincidence but i am trying to create a fight scene were someone punches another person but over half the Images have the wrong person punching. and the ones that have the right person doing the punching usually have another issue like the usual messed up hands or other fuckups. i made something similar a few weeks ago and it went perfectly on the first try

pixxelkick@lemmy.world · -1 pts · 1y (3 replies)

I would say your best bet is typically using inpainting.

Start with just getting all the figures framed/positioned well, such that they look about right in terms of stance and position in the frame.

At this stage only use descriptors that describe all 3 of them well.

Then inpaint only one of them, and add in all the descriptors just for that figure, and regenerate just that section with a bit higher noise level until she looks right.

Repeat for each one individually.

GrumblePuss@lemmy.world · 1 pts · 1y

Perchance image gen does not support inpainting.

Jack_Fosset@lemmy.world · 1 pts · 1y (1 reply)

In my experience inpainting looks like a part from a different image was snapped in and doesn’t blend well with the original in most of the cases

pixxelkick@lemmy.world · 1 pts · 1y

You need to use a model trained for inpainting.