i think you can only do Textual inversion with 6 gb
i think you can only do Textual inversion with 6 gb
Flora still requires 8
HYes Embeddings is really the main thing I use in my workflows right now
SHi everyone, in you experience, is a larger training set better than a smaller one in dreambooth? Like do i get better results with 10 pictures of myself or 100? Or do I just get overtraining?
Fjust have high quality pictures
SGot it, thx!
Fi think today i will make a tutorial for deep-floyd/IF
Four subscriber rate is still 10% i dont get why :/
AHi, I'm trying to train a LORA model in Dreambooth on Automatic1111. I got it to kind of work a few times but now I can't seem to get good results. It seems like using a class data set along with my images makes the result worse. If I share my parameters/data info could someone give me some tips on what to change?
Rdoes anyone know how to update the Lora models in Auto1111 when you click refresh. I actually have to quite the terminal and relaunch from scratch. Refresh Lora tab and reload webui, neither refresh the new model in. Any thoughts?
Fhi welcome. i am waiting sd web ui update to make a new lora video but sure you can share settings
Fit dont see models?
AI just got some pretty good results with this data and these settings:
42 training images of myself in different situations/lighting
caption .txt files that go along with the images
RealisticVisionV20 as a source checkpoint
Saving
-Use Lora
-use Lora extended
-100 steps/epochs
-batch size 2
-use gradient checkpointing
-0.00001 unet learning rate
-constant_with_warmup LR scheduler
-other settings from all the vids, 8bit AdamW, fp16, xformers
-Scale prior loss (checked, did not edit values)
-no sanity prompt
Concepts
-Used directory with photos of me, and .txt files with the token I assigned replacing my name
-no class data
-instance token is my name without vowels
-class token is man
-instance and class prompts are [filewords]
-sample prompt boxes are blank
-no class img gen
I'd rate the results i'm getting 7 or 8 out of 10, any tips appreciated
42 training images of myself in different situations/lighting
caption .txt files that go along with the images
RealisticVisionV20 as a source checkpoint
Saving
-Use Lora
-use Lora extended
-100 steps/epochs
-batch size 2
-use gradient checkpointing
-0.00001 unet learning rate
-constant_with_warmup LR scheduler
-other settings from all the vids, 8bit AdamW, fp16, xformers
-Scale prior loss (checked, did not edit values)
-no sanity prompt
Concepts
-Used directory with photos of me, and .txt files with the token I assigned replacing my name
-no class data
-instance token is my name without vowels
-class token is man
-instance and class prompts are [filewords]
-sample prompt boxes are blank
-no class img gen
I'd rate the results i'm getting 7 or 8 out of 10, any tips appreciated
Fnice thanks for letting us know
Fhave you compared with kohya lora?
AI haven't
Ais it a lot of work to setup? Figuring out SD and Automatic1111 and Dreambooth has been a lot lol
Frecording a video for DeepFloyd-IF
Ayour videos have helped me a lot already, looking forward to your future uploads!
Fthank you so much
RHi @Furkan Gözükara SECourses got a question. Trying to update StableTuner to Torch2.0. I am looking at your tut. Yours is venv/Scripts/activate... But i installed virtual environment with miniconda3. So would i do the same in envs/ST/Scripts/activate (ST is the name of the environment).. everything else the same?
Rpip install --force-reinstall torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install --force-reinstall --no-deps --pre xformers
pip install --force-reinstall --no-deps --pre xformers
Adoes anybody know how to turn a dreambooth model into an inpaint model?
Aa atempted to use merge checkpoint and use an inpaint model with the dreambooth model and then put vanilla 1.5 as C and do add difference, but it completly messes up the dreambooth model and makes the trained subject barley useable..
Fi think it should be
Fdon't know :/ have you asked on reddit?
Rhey, i dont go on reddit much. what channel on reddit should i go to?
RBy the way, did you know you can use Stable Tuner's api to monitor epoch training via Telegram? Its pretty cool
Fmain stable diffusion
Fwow nice
ROh right, cool will check
Rso you can live updates every time the training spits out samples for each epoch
Fdo you want me to test any prompts with DeepFloyd-IF
Fdmn their naming so stupid
Rtry something like "photo of fashion model walking in the street in london in the style of i-D magazine 2000's by mario testino, masterpiece, film camera
Dhello, would you guys know what could cause kohya lora training not to work?
Dcouldn't fine anything online
Ayea,, there is this guide: https://www.reddit.com/r/StableDiffusion/comments/zyi24j/how_to_turn_any_model_into_an_inpainting_model/
but the problem is that it doesn't take dreambooth in to account.. I'm sure it works on ordinary models but for dreambooth it seems to completly mess up the trained part of the dreambooth model, almost as if it's too much of a specific part of the model, where ordinary models have a more general training data..
but the problem is that it doesn't take dreambooth in to account.. I'm sure it works on ordinary models but for dreambooth it seems to completly mess up the trained part of the dreambooth model, almost as if it's too much of a specific part of the model, where ordinary models have a more general training data..
reddit
322 votes and 93 comments so far on Reddit
Adid you train a locon ? or ordinary lora?
Djust Lora I don't even know what locon is
and it didn't even strat training
and it didn't even strat training
Dit just gives out some vague error
Athis is the problem.. what model are you trying to train on ?
Fon which model you are doing training?
Drealistic vision 1.3 I think it was but I tried the normal SD 1.5 and same thing happened I don't think the model is the problem
Fcan you try without mem eff attention
Falthough it indicates model structure error
Fi suggest to first try with some of the base models
Dokay I will right now