#QuickStart sd1.5webui + dreambooth + training tools

30 messages · Page 1 of 1 (latest)

delicate chasm
#

Link: https://runpod.io/gsc?template=19of8e9n4i

Here you can watch an example of how to use it: https://www.youtube.com/watch?v=TWzL-vQM7vk

UPDATE: March-25 - VAE and 1.5vae checkpoint included

I created this template with webui and a set of opinionated (working) extensions to assist in training and refining checkpoints and embeddings.

Launches super quick as everything is pre-installed on the Docker image.

Optimized to run on 3080s and up (works great on the 3090s and A4000s).

Includes:
WebUI
Tensorboard
CUDA 11.8
OpenCV
1.5_pruned_emaonly (no download on startup)
VAE

Extensions

  • Dreambooth (working fine)
  • DreamTrain 2
  • Dynamic Prompts/Promptgen
  • DAAM
  • Deforum
  • Inspiration
  • Embedding editor
  • Training+
  • Sane defaults
  • More useful tools...

I hope this is useful for you.
Let me know if it works for you, and/or if you would like to see something added on a future version.

GH repo: https://github.com/bbaaxx/runpod-sd15drb

glacial carbon
#

Hey that's awesome! Thanks for making this 🙂

pale cipher
#

hey man do you have guide for training?

#

@delicate chasm ??

delicate chasm
#

@pale cipher I have not authored one yet, but you can find good guidance from AITREPRENEUR videos here. The tools he uses are all on the template.
https://youtu.be/2ityl_dNRNw

trail bear
delicate chasm
#

@trail bear A password is randomly generated and is stored on an ENV variable on your pod settings (under edit pod).
You can use that one or change it from there.

restive shuttle
barren halo
#

Thank you so much for putting this together! I feel bad asking you directly so... I'd like to see if anyone has had any luck with the Dreambooth extension on this template.

#

I'm seeing in logs: No module named 'dreambooth.dreambooth'

delicate chasm
#

The latest versions of Dreambooth change the requirements a bit which is causing the base setup to break I will make an update of the image today. (basically you need Python3.10.9)
Sorry for the inconvenience. @barren halo and @restive shuttle

#

@restive shuttle Port 4204 is because during trining/dataset prep sometimes you need to launch a secondary tool such as lama-cleaner or ComfyUi and due to the way Runpod works you have to restart the container just to open a new port (you lose configs and things like that). So I leave that there as a utility

delicate chasm
#

@barren halo I pushed an updated version of the Docker image (with dreambooth updated to latest version). Have not tested it yet on Runpod but "it ran on local =)". If you care to give it a go.

#

There is an alternate image with both VAEs (one on the 1.5 and the added newest VAE.safetensors) "bbaaxx/sd15db-runpod:v15-vae"

delicate chasm
#

I have updated this to support torch 2 and xformers(pre), now is more stable and slightly faster on big GPUs

dawn mountain
delicate chasm
#

@dawn mountain I just installed it by compiling Python 3.10.9 first ... but I just discovered that Dreambooth does not work with the required accelerate version to run on torch 2 so for now I will revert it to 1.13 until proper support is added 😥 Everything else seems to work just fine.

tawdry holly
#

@delicate chasm I just discovered RunPod SD template has torch 2 as default, but I need to downgrade to torch 1.13 for Dreambooth. I don't have much coding experience. Would you kindly point me to resources that I can learn how to do this? Also, would this require Docker knowledge? Thank you

delicate chasm
#

There are a few ways you can do this:

  1. Use an older revision of the runpod container (I think version 4 should have torch 1)
  2. Downgrade the version of Webui you are using via git pull command (specifying which commit to go back to)
  3. Manually downgrade Torch (using pip install command)

I suggest for dreambooth to use option 1 or 2.

No need to use docker for a one-time torch change but if you plan to use it often, then probably you will want to make your own Docker image (quite involved)

#

Alternatively, you can look into the Khoya-SS GUI or the khoya dreambooth notebooks as they seem to be a better alternative to Dreambooth

tawdry holly
#

@delicate chasm thanks so much. I was about to manually downgrade Torch, and found my pod has been shutdown for "emergency repair". haha... In any case, I launched your template instead in a new pod. Would you know why I cannot launch webui at Port 4206? I tried so many times and keep getting the bad gateway

tawdry holly
#

@delicate chasm I'm not having luck launching at 3 new pods. I will try again later, but I'm wondering if port is in use?

tawdry holly
#

@delicate chasm sorry to spam your message board, just want to give you an update. Finally I was able to launch to port. Now the problem is port 4206 is in use. I tried setting 4 differnt ports, no luck. Each port seems to be in use. Do you know what is the issue I'm experiencing?

delicate chasm
#

Uhmm, that is weird. I will take a look but the webui should be running on :4206. Try just rebooting the pod

delicate chasm
#

@tawdry holly I just tested the container on an A4000 and it seems to be working fine on port 4206. The relauncher.py file sometimes will not let gradio free up the port entering in a kind of loop.

  • You can either restart the pod (Recommended, as it should recover on its own).
  • or add the environment variable: RUNPOD_STOP_AUTO=true to prevent the script from launching the WebUI and then you can launch it manually from a command line with the following commands:
source /workspace/venv/bin/activate
cd /workspace/stable-diffusion-webui-ux
python relauncher.py

for example.

However my base image has Torch 1.13.1+cu117 and xformers 0.0.16rc425.
Dreambooth seems to be working fine (screen attached)

Feel free to fill up an issue on the GitHub repository if you get an error (I will try to take a stab at the relauncher.py issue on a future release)

tawdry holly
tawdry holly
#

@delicate chasm Beautiful 😽 it's up and running. I think maybe this morning RunPod was having issues. I'm on A5000, so far so good. Thanks for making the template.

woeful widget
#

I can't get passed this error. Tried restarting the pod, starting on a new pod, etc.
2023-05-30T02:01:50Z error creating container: Error response from daemon: failed to mkdir /var/lib/docker/volumes/mmd5wvr874kjii/_data/stable-diffusion-webui-ux: mkdir /var/lib/docker/volumes/mmd5wvr874kjii/_data/stable-diffusion-webui-ux: disk quota exceeded

jade crater
regal elm
#

Not interested in OpenPose?