kohya_ss
StableTuner
kohya_ss | StableTuner | |
---|---|---|
132 | 22 | |
8,414 | 626 | |
- | - | |
9.9 | 10.0 | |
3 days ago | about 1 year ago | |
Python | Python | |
Apache License 2.0 | GNU Affero General Public License v3.0 |
Stars - the number of stars that a project has on GitHub. Growth - month over month growth in stars.
Activity is a relative number indicating how actively a project is being developed. Recent commits have higher weight than older ones.
For example, an activity of 9.0 indicates that a project is amongst the top 10% of the most actively developed projects that we are tracking.
kohya_ss
-
Some semi-advanced LoRA & kohya_ss questions
Many of the options are explained here https://github.com/bmaltais/kohya_ss/wiki/LoRA-training-parameters
-
Lora training with Kohya issue
training in BF16 might solve this issue from what I saw in this ticket. I know other people ran into the issue too https://github.com/bmaltais/kohya_ss/issues/1382
-
What is the best way to merge multiple loras in to one model?
for lycoris loras you can use the command-line script from the kohya-ss repo https://github.com/bmaltais/kohya_ss/blob/master/networks/merge_lora.py i have an older version checked out from late july, it had a separate merge_lycoris.py for for this purpose, it's probably unified now in a single file
- Evidence that LoRA extraction in Kohya is broken?
-
Merging Lora with Checkpoint Model?
I usually do that with kohya_ss, a tool made for making LoRAs and finetunes. It might be a bit of a pain to set up just to do this one task, but if nobody gives you an easier method, look into it. https://github.com/bmaltais/kohya_ss
-
How I got Kohya_SS working on Arch Linux, including an up-to-date pip requirements file
After that, make your staging directory, and do the git clone https://github.com/bmaltais/kohya_ss.git, and navigate inside of it. Now, here's where things can become a pain. I used pyenv to set my system level python to 3.10.6 with pyenv global 3.10.6, though you can probably just use "local" and do it for the current shell. You NEED it to be active however before you set up your venv. If you do python --version and get 3.10.6, you're ready for this next part. Make your venv with python -m venv venv. This is the simplest way, it'll create a virtual environment in your current folder named venv. You'll do a source venv/bin/activate and then do which python to make sure it's using the python from the venv. Now for the fun part. The included setup scripts have been flaky for me, so I just went through the requirements and installed everything by hand. I'm going to do this guide right now for nvidia, because I just got a 4090 for this stuff. If this ends up working well for others and there's demand, I'll try to reproduce this for AMD (But I'll be honest, I got an nvidia card because bitsandbytes doesn't have full rocm support, nor do most libraries, so it's not very reliable). After installing everything and testing it works at least at a basic level for dreambooth training, my finished requirements.txt for pip is as below:
-
The best open source LoRA model training tools
Earlier I created a post where I asked for recommendations for LoRA model training tutorials. The first one I looked at used the kohya_ss GUI. That GitHub repo already has two tutorials, which are quite good, so I ended up using those:
-
Script does...nothing
I have tried my best to research this issue and have not come up with much. It is obvious that its a backend issue right? The guides that I used https://github.com/bmaltais/kohya_ss and https://github.com/pyenv-win/pyenv-win/
- Using LoRa on SDXL 1.0 (not using the Kohra GUIs)
-
How do I reduce the size of my Lora models?
I am training on a 12GB 3060 using kohya_ss. Is there a setting or something I'm missing?
StableTuner
- What is the best way to train a Stable Diffusion model on a huge dataset?
- How to fine-tune a Stable Diffusion model with hundreds or thousands of images?
- SD fine-tuning methods compared: a benchmark
-
After so many errors with Dreambooth, Everydream2 is the way to go
Of all dreamboothing/finetuning implementations I tried I liked StableTuner the most. Might be worth giving it a shot to compare as well.
-
Non-technical tips for ideal training of Stable Diffusion through Dreambooth?
Largest I've gone is about 100 images for objects or people. I don't think it matters though, it can be a hassle setting up and resuming the training session each time if your doing small sessions. Stable Tuner can simplify all of this by helping you set everything up through their client installed locally. You can then easily do your training locally in short sessions or have it automatically packed up to be exported to colab or another gpu hosting service, also with the ability to train in short sessions. Its a smart way to manage large training projects like yours. It requires a bit of time setting up but most folks who have already played around with dreambooth should be able to navigate their way through easily enough. It has all the other training methods built into it too, including proper fine tuning https://github.com/devilismyfriend/StableTuner
-
Alternative tools to fine tune stable diffusion models?
Some people also like StableTuner: https://github.com/devilismyfriend/StableTuner
- Question about specific character training
-
Finetuning Inpainting model
Stable Tuner seems like it's setup to allow training on regular/inpaint/depth models. https://github.com/devilismyfriend/StableTuner
-
The next best alternative to Auto1111??
StableTuner is an alternative to the sd_dreambooth plugin. It can do Dreambooth and Fine Tuning (I haven't tried this but I think it's embeddings) It uses diffusers but will convert between that and ckpt files, is for Windows/Nvidia, and uses a local app instead of a webapp. This is the only successful local Dreambooth I've done. You'll need to go to their discord for help but it's not hard to use.
-
Auto1111 Fork with pix2pix
Dreambooth is has better results in older commits. StableTuner is better for training : https://github.com/devilismyfriend/StableTuner
What are some alternatives?
sd_dreambooth_extension
EveryDream2trainer
EveryDream-trainer - General fine tuning for Stable Diffusion
Dreambooth-Stable-Diffusion - Implementation of Dreambooth (https://arxiv.org/abs/2208.12242) with Stable Diffusion
sd-scripts
LyCORIS - Lora beYond Conventional methods, Other Rank adaptation Implementations for Stable diffusion.
automatic - SD.Next: Advanced Implementation of Stable Diffusion and other Diffusion-based generative image models
ComfyUI - The most powerful and modular stable diffusion GUI, api and backend with a graph/nodes interface.
kohya_ss_colab - a (successful) attepmt to port kohya_ss to colab
LoRA_Easy_Training_Scripts - A UI made in Pyside6 to make training LoRA/LoCon and other LoRA type models in sd-scripts easy
dreambooth-training-guide