I intend to use a 960 x 512 when working. On the preprocess images it says 512x512
I intend to use a 960 x 512 when working. On the preprocess images it says 512x512
Fyour classification images probably should be same dimension
Fas your training images
Fif you dont change max resolution they will get downscaled
Ygreat, thanks a lot!
Ykeep up the great work
YIt worked! im so happy!
RHey @Furkan Gözükara SECourses are you aware of any GPT assistants who can already help with captioning datasets. I find myself constantly going back to make changes to a tag word or a string, and have to go back to all the dataset txt files, and it's quite time inefficient. Would imagine being able to leverage GPT + python would bring captioning from an hours to minutes. Any luck yet with applying these GPT agents to captioning?
Fthere are already blip and other captions
Fwith the caption thing what is important is not describing that image
Fthe important thing is which words you want to improve
Fyou want to do fine tuning right?
Dalthough BLIP makes the process much faster, in my opinion it is less accurate. Sometime it produces wrong descriptions for the images. This might affect the training right? I have not tried GPT-4's image description feature yet, but from what I heard, it is quite accurate. maybe a comparison between BLIP and GPT-4 is needed
RYes precisely. I am doing Lora of recognisable fashion model person wearing recognisable fashion outfit. I need to be able to trigger the recognisable face and clothes, so it is rigid, and to do that I am using key tags in the caption to make the ai aware what is in the picture. I havent optimised yet for face, i am working on outfit to be faithfull to my studio original photo reference that I shot. But the point is, that BLIP captioning only creates the txt structure. The caption guess is pretty useless. But anyway, what i want is an Ai assistant to propagate the editing. So i decide, i need to change a term, say about the buttons, i need the button tags to change across all the txt files in the dataset. Do you know what I mean.
RI am just asking about an Ai assistant to go in to each txt file and change them. I am using 50 images per dataset concept, and i have about 10 concepts (outfits). So that's a lot of txt files to recursively edit.
RSo I would imagine a python code would be involved to delegate text editing across all text files...
RI want to fine tune yes.
Dso you want to replace a tag with another tag and do the same for all the text files?
RYes. Say, "white button" to "4 white patterned button". There are 50 images with those buttons. Would be great for the Ai assistant, to propagate the change across all the txt files
RCLIP captioning only guesses what things are. I need something to help me edit details
RSo yes, replace tags
R--across a group of txt files
Dyou could download notepad++ and use "find and replace in files" feature, it will find and replace words in all the files you have mentioned
Roh i did not know that. It can do this already
RHow silly of me
ROk, will give it a try now
Dyes, I learned this 2-3 days ago as well, maybe from @Furkan Gözükara SECourses videos or from other dreambooth tutorials on YT
RAs an extension though, It would be ideal if an Ai assistant could be fine tuned to execute things in my captioning categories. I always say, "head turned to side" or "head tilted". etc. Wouldnt an Ai assistant be ideal to learn what you like to caption and do it for you? But i suppose this is still some ways off
Dctrl+f > find in files > "input directory where files are present", in Filters put: ".txt"
D
Dyes, this could be possible
Rmaybe with MiniGPT?
Dohh, miniGPT do have Image describing feature
KI'm using Runpod
KDoes anyone have a problem where Dreambooth Load Setting doesn't load and keeps loading forever? I tried to load a 100-epoch checkpoint to train it until 150 epochs, but it keeps loading. I have faced this problem several times in the past, but I wasn't sure how to fix it. I have restarted SD multiple times sometimes the error goes away on its own, sometimes it doesn't.
Fok you want to fine tune which words?
Fit shouldnt take that long
KYes I don't know why but sometimes it keep loading forever
FAre you connecting via http 3000 port?
KYes
FNot much else you can do
RI am using Kohya_ss Dreambooth/Lora module. I shot my dataset in studio. I feed around 50 images. The purpose is maintain rigidity in the output model, so the concept trigger (this real person + this brand outfit, long coat with buttons and pockets, ruffled collar silk shirt, brown ankle boots). I need the Lora model to spit out precisely the same character concept every time, in different environment settings. So my captions will differ in each cropped dataset. FOr example, for close-up: lena, woman, fashion model, fair skinned, skinhead, short hair, white stripped ruffled collar shirt, buttons, grey coat with white speckles, head and shoulder close up, head turned to side, brown eyes, beige background . For full lenght wide shot: lena, woman, fashion model, standing, fair skinned, skinhead, short hair, white stripped ruffled collar shirt, buttons, grey coat with white speckles, white coat buttons, coat pockets, coat notch lapel, brown leather ankle boots white stripped sleeves, brown eyes, beige background . I am being precise with tags, so that i get an identical generation of the concept character. Fine tuning the captions for me gives me better precision of garments. I am getting pretty good results, but fine tuning caption tags means going back and recursively changing tags (ie, "white stripped ruffled collar shirt" to "white stripped silk (ruffled collar:1.2) shirt, long sleeve"... etc). That's why I ask about a tool automate the propagation of tags...
Rso presumably, one could train the ai there to understand what details i want to automate, ie [head tilting | turning to side | facing_camera]... I wish someone was already doing this with miniGPT
Rlet me try this
DminiGPT also needs 12GB VRAM minimum, that is a disadvantage as well.
乂Hi, I am getting this error "'A' is not recognized as an internal or external command, operable program or batch file" while installing Kohya_SS.
乂I have checked and explored it further. The error is being caused when setup.bat executes the venv/scripts/activate.bat
Fprobably you need to code an application
Fthat can do that
Rok i see. so probably not for me, since i am no coder
Ri have a 4090