Could you put a way for us to do manual captions on the training images to get better results than clip captioning?

[artificialguybr]

Title

[artificialguybr]

https://github.com/TheLastBen/fast-stable-diffusion/discussions/427

At first the answer to my question is here. As I understand it, it is the same thing, right?

[TheLastBen]

you can caption each image if you want, you don't even have to use "_" to separate the words, you can name the files in the format : "a man driving a car.jpg"

[artificialguybr]

you can caption each image if you want, you don't even have to use "_" to separate the words, you can name the files in the format : "a man driving a car.jpg"

Do I have to put the instance name in the middle and it will recognize it the same way?

like instance name mafk4

mafka4 in the middle of the jungle.png

mafka4 smiling.png

mafka4 crying.png

I don't need to put the numbers in the images too? like (1) (2) (3)

[TheLastBen]

you don't need to put the numbers if the names are different, the instance name must be present in all images, it doesn't matter the place, and avoid using numbers in instance names, they will be automatically removed, your instance name would become : mafka instead of mafka4

[LIQUIDMIND111]

you don't need to put the numbers if the names are different, the instance name must be present in all images, it doesn't matter the place, and avoid using numbers in instance names, they will be automatically removed, your instance name would become : mafka instead of mafka4

can you use CAPITAL letters? like maFKa smiling.png or mafka smiling on THE PARK.png

[TheLastBen]

sure, you can

[LIQUIDMIND111]

sure, you can

hello ben, this is rare, i am using the old model AS ALWAYS, i have done more than 50 models in 2 months with your OLD model page, and last night i tried to make a model of a friend, and it comes out very badly, does not look like her, i meant, the model is HD, very well-trained, but DOES NOT look like her.... i tried WITH or without text encoder, and i even tried to do AGAIN a past model using the SAME source images and caption instance, and DONT get the same result.... did something happened to the codes or something that i need to know? her instance is GFSDGANGxIE and if you add sexy before GFSDGANGxIE it will turn the model into a chinese woman.. lol rare....

[TheLastBen]

Hi, I didn't change the old method notebook recently, so it just might be the instance images, and try setting the text encoder to 100%

[LIQUIDMIND111]

Hi, I didn't change the old method notebook recently, so it just might be the instance images, and try setting the text encoder to 100%

ok let me clarify, the page works PERFECT without the IMAGE CAPTION THING, a week ago it was fine, but today, iget NICE HD models, but are not either me or my friend..... and i even fixed the instance names without NUMBERS as suggested..... and now i get rare people....... many chinese woman, it slike its showing more like the CLASSES people that i used, rather than the SUBJECT photos..... but i wil try the text encoder at 100

[artificialguybr]

When I tried to put the caption on the images, it ended up recognizing each image as a different instance.

Examples

a_dolphin_madeof wood_madeistily.png

a_dog_made_of_wood_madeistily.png

a_hippopotamus_made_of_wood_madeistily.png

I made something wrong?