@Dr. Furkan Gözükara if I train for 1024x1024 resolution, can I also provide 2048x2048 or do you alw
@Dr. Furkan Gözükara if I train for 1024x1024 resolution, can I also provide 2048x2048 or do you always recommend 1024x1024
FDid you have any success with multi_gpu training?
For me, In general, it seems to attempt launching training on multiple GPUs but then I get an out of memory error which terminates everything.
Training on 1 GPU with all the settings the same works. So it seems multiple GPU somewhere need more vmem on a single card
For me, In general, it seems to attempt launching training on multiple GPUs but then I get an out of memory error which terminates everything.
Training on 1 GPU with all the settings the same works. So it seems multiple GPU somewhere need more vmem on a single card
F1024x1024 i suggest
Fmaybe like 1024x1280
Fyes on kaggle i trained
Fit trains with dual gpu
SI will give it a try. to run it on runpod is with fast stable difussion? works with automatic 1111?
Fit is working with any since it makes own unique venv
Fi am trying to fix an error
Fon rupod
Fye we need to use
Fwith gradio share
Fon runpod
Swhat template should I pick on runpod? when you finish the repair?
Sthe same process as the interrogator?
Fany template . i selected same
Frunpod/pytorch
Fyes share working
F
Fneed gradio 4.0 for local proxy
F
FI have access to 8x 3090 systems and can not get it to work 
Such a shame as I have all this processing power
I'll compare the Kaggle tutorial step by step with my setup
Such a shame as I have all this processing power
I'll compare the Kaggle tutorial step by step with my setup
Sthis is midjourney killer
Fi can make it work
Sno commercial license to pay?
Fwith private consultation
Flet me look licence
FGitHub
Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis - PixArt-alpha/PixArt-alpha
Sit seems that is free commercial use
S
Famazing
Sno training on this new model yet right?
DThanks to @Dr. Furkan Gözükara , Kaggle (for this wonderful free platform) , Hugginface( for huge space and repository)
Donce this project is over , next will be animatediff (hopefully you come out with similar guide for us sooon :D)
Djust found lower numbered lora dont even take the prompts at all..it just isnt trained or trained too much on the training images..
AWill you have a tutorial on using esrgan models @Dr. Furkan Gözükara ?
Dnot having much fun with dreambooth checkpoints .. 
GN anyway...
GN anyway...
NHi everyone, what I noticed is that my CN on SD RunPod is this version: v1.1.313, and it doesn't update to the higher version.
NI've tried everthing from tutorials but didn't work
NI am using RunPod's Template
JStable Diffusion XL (SDXL) DreamBooth Training eror (((
JFile "/opt/conda/bin/accelerate", line 8, in <module>
sys.exit(main())
File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/accelerate_cli.py", line 47, in main
args.func(args)
File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/launch.py", line 977, in launch_command
multi_gpu_launcher(args)
File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/launch.py", line 646, in multi_gpu_launcher
distrib_run.run(args)
File "/opt/conda/lib/python3.10/site-packages/torch/distributed/run.py", line 785, in run
elastic_launch(
File "/opt/conda/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 134, in call
return launch_agent(self._config, self._entrypoint, list(args))
File "/opt/conda/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 250, in launch_agent
raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
=====================================================
./sdxl_train.py FAILED
-----------------------------------------------------
Failures:
<NO_OTHER_FAILURES>
-----------------------------------------------------
Root Cause (first observed failure):
[0]:
time : 2023-11-27_18:01:49
host : c4565fefbe4d
rank : 1 (local_rank: 1)
exitcode : -6 (pid: 1053)
error_file: <N/A>
traceback : Signal 6 (SIGABRT) received by PID 1053
=====================================================
sys.exit(main())
File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/accelerate_cli.py", line 47, in main
args.func(args)
File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/launch.py", line 977, in launch_command
multi_gpu_launcher(args)
File "/opt/conda/lib/python3.10/site-packages/accelerate/commands/launch.py", line 646, in multi_gpu_launcher
distrib_run.run(args)
File "/opt/conda/lib/python3.10/site-packages/torch/distributed/run.py", line 785, in run
elastic_launch(
File "/opt/conda/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 134, in call
return launch_agent(self._config, self._entrypoint, list(args))
File "/opt/conda/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 250, in launch_agent
raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
=====================================================
./sdxl_train.py FAILED
-----------------------------------------------------
Failures:
<NO_OTHER_FAILURES>
-----------------------------------------------------
Root Cause (first observed failure):
[0]:
time : 2023-11-27_18:01:49
host : c4565fefbe4d
rank : 1 (local_rank: 1)
exitcode : -6 (pid: 1053)
error_file: <N/A>
traceback : Signal 6 (SIGABRT) received by PID 1053
=====================================================
MI discover the hidden things of Prodigy. For example, if I don't have 25 pictures, but 21 or 26, I get more exciting results. For example, with 26 images, I had to lower the d_coef from 1.02 to 0.85, and even then Lora was a bit strong.
Even at 21 images it might be too strong if I set it to 4 repetitions instead of 5. So there are still some small things to explore, but the fact is that whether Lora is stronger or weaker, it still gives a nice result, as long as I fine tune its flexibility.
Even at 21 images it might be too strong if I set it to 4 repetitions instead of 5. So there are still some small things to explore, but the fact is that whether Lora is stronger or weaker, it still gives a nice result, as long as I fine tune its flexibility.
MHere is a sample from SD 1.5 with 768x768 and without Hires Fix
MI trained model with 512x512 images
TDoes anyone know the formula for number or regularization images? Something like repeats * train image count / batch size == reg images?
JI can't use more than 13 training photos. error
TIncluding the error in your message might help
Fthey have script but 0 info
Fgood job
Fyou just put them into folder test