can we run deep floyd locally on 11 gigs of vram?
can we run deep floyd locally on 11 gigs of vram?
Hi tried not using classification dataset and class image name in lora, the result is not bad
H120 instance images
Hwhat is the optimal value of the train unet (steps ratio of the textencoder ) for train a style /person
P@Furkan Gözükara SECourses I followed your tutorial to install ControlNet (https://www.youtube.com/watch?v=YJebdQ30UZQ). When I get to the step where I run python gradio_scribble2image.py for the first time, an error message is returned:
RuntimeError: CUDA out of memory. Tried to allocate 2.00 MiB (GPU 0; 4.00 GiB total capacity; 3.19 GiB already allocated; 1.40 MiB free; 3.25 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Do you know what I'm supposed to do to fix this error ? thks a lot !
RuntimeError: CUDA out of memory. Tried to allocate 2.00 MiB (GPU 0; 4.00 GiB total capacity; 3.19 GiB already allocated; 1.40 MiB free; 3.25 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Do you know what I'm supposed to do to fix this error ? thks a lot !
Fnope. i am showing on free kaggle
F30 hours gpu per week
Fit is said to be 75% but i didnt test extensively
Ffor style it is said to be 25% but i used 50%
Fprobably doesnt work with 4 gb. maybe automatic1111 optimizations can run it
FYouTubeSECourses
Our Discord : https://discord.gg/HbqgGaZVmr. In this video, You will learn how to use new amazing Stable Diffusion technology #ControlNet in Automatic1111 Web UI. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on 
https://www.patreon.com/SECourses
Playlist of #StableDi...
https://www.patreon.com/SECoursesPlaylist of #StableDi...
Fi am hopefully also planning a newer video too for controlnet 1.1
Fstill should work
Pthank you, i'll try
Zhey @Furkan Gözükara SECourses I'm trying to use your regularization images for kohya, but everytime i do, it says 0 reg images during training, although it says 640 train images with repeating
Znvmd figured it out
RI just realised, that I installed Kohya_ss choosing Torch 2.0, and you mentioned in the video that CuDNN is not necessary with Torch 2.0. So that installing it is not necessary. So case closed. Good thing I watch your LoRA video to catch that detail. Thanks again
HGitHub
Stable diffusion webui based on diffusers. Contribute to ddPn08/Radiata development by creating an account on GitHub.
H@Furkan Gözükara SECourses
Z@hopeyan does it run better than automatic1111?
RYou still like webui dreambooth better than koya/Lora? Your video on Kohya Lora looks really good
Zdeepfloyd has to cut down on it's vram usage, 16gb is kinda ridiculous for making some text
МIt turns out that you can enter an arbitrarily long request into a negative prompt and its accuracy will not decrease?
Fyou are welcome. yes torch 2 has 8.7 dlls
Fvery nice supports DeepFloyd IF
Fi should make a tutorial for this
Factually i am comparing atm :d
Fdoing rank 128 192 and 256 lora training
Fi have done dreambooth
Fwill compare
Fye even lowest is high
Fbut in video i used kaggle
Fso everyone will be able to use
Fno it will. negative prompts also do affect your final results significantly
Ddr. F did u ever figure out the batch loader for control nets?
Das in dif batches for each control net seperately
Fyes there was a pull request and it had worked for me,
Fwow lora super memorized the background and clothing interestingly
Fi will test with captions
Fdreambooth had much better flexibility
Zdid u upload the video yet? never heard of kaggle before
Zyeah i've had much better success with captions, without, my model was burn after 2 epochs
Fstill processing
Fkaggle is even better than colab
Fye i am testing git captions
Fhere lora vs dreambooth
Zi usually do mine like this
Fi will show in a minute
Z
Zafter i do blip captions, run the following command - sed -i 's/a man/jrch/g' *.txt and replace man with woman or whatever i want, and jrch with ohwx
Rwhat are you using here?