I didn't understand how to make it work.

[hseapi]

I may not be being clever, but I couldn't figure out how it works. I couldn't find the part that puts the images, captions and regularization I want. And I don't think I'm inexperienced. If it is not too much to ask, It would be nice a summarized tutorial, it would be of great help.

[Nerogar]

You can find a short guide here. https://github.com/Nerogar/OneTrainer/blob/master/docs/QuickStartGuide.md (Also linked in the readme file)

[gsusgg]

@hseapi If I'm understanding you correctly, its the concepts tab, add concept, then just click on the new concept that showed up.

[shadowlocked]

I have to agree that while the GUI is nicely laid out, it is massively unintuitive to use, particularly for people coming from Kohya - who, for instance, are used to the (almost equally clunky) method of setting the trigger word in the folder name of the source images. Both Kohya and OneTrainer need a simple text field where you can type in 'TomCruise', 'DarthVader', or whatever. It's incredibly difficult to figure out. There are no tooltips (despite what's said in a Reddit post announcing it, they never appear for me, and it's not like I'm running it in the wrong browser or something), the Discord is under-populated and the Wiki has many omissions, examples of strange or elliptical phrasing, and just in general assumes massive prior knowledge of a really eccentric and unique schema.

Unless someone stops coding for a little while and starts documenting, I think OneTrainer is going to continue to preach to a pretty small crowd of converts, cognoscenti and an 'inner circle'. And it's a shame, because it has some potential, and offers masking, and many other innovations.

We don't need a Clippy wizard - just some tooltips and a cohesive Wiki that answers questions that (if you look in the Discord) have some really unexpected answers (i.e., the 'trigger word' is the name of the model file, even though there are about four other possible places, referring to 'prompts', that the average user might suspect this to be located, in the Concept widget).

[KallutoZoldyck]

As a Kohya user myself, who is relatively new to this space, I've realized that OneTrainer was actually pretty easy to figure out within my first few minutes of looking at the ui. I've found that there are tons of tools in it that speed up the preparation process, and I recognize that it has the potential to be the most intuitive trainer, and also yield the best results due to the mask training. While it's a shame that the community is small at the moment, I'm certainly going to take advantage of that by writing them my list of suggestions, and learning what works with it, and what doesn't. Overall, this tool was very easy to figure out, and it had instructions on the github page that I read through to help figure everything out.

[gabeacid]

there are about four other possible places, referring to 'prompts'

Note for embedding you need to include <embedding> in the prompt, not the embedding name but just <embedding> that acts as a placeholder. Either in the image caption or in the single text file.

And in the concept window there is an input labelled Name.

For LORA:

• QuickStart: " In that window you can optionally select a name, then specify the directory where the images are saved. The prompt source defines where the prompts are loaded from"
• Wiki: "Note for embedding you need to include <embedding> in the prompt, not the embedding name but just <embedding> that acts as a placeholder. Either in the image caption or in the single text file."

The embedding template: a photo of <embedding>
But the generated caption files: a pthoto of a man|woman|etc

Dataset prep tool (blip) didn't put the placeholder into the caption files, so which format is proper:

1. a photo of a man|woman|etc
2. a photo of a <embedding> man|woman|etc
3. a photo of a <embedding>
4. <embedding>

If the concept name (in the concept window) and the dataset folder name irrelevant and the embed|caption prompt needs to have <embedding> not embedding name how to control multiple concepts?

1. datasets_folder\car_folder\caption_xxx.txt: <embedding> car
2. datasets_folder\bus_folder\caption_xxx.txt: <embedding> bus

Validation prompts need to be contain the lora name? If yes what format?
<embedding> or EMBEDDING_FILE_NAME?

1. a photo of a man|woman|etc
2. EMBEDDING_FILE_NAME a photo of a man|woman|etc
3. a photo of a EMBEDDING_FILE_NAME
4. a photo of a <EMBEDDING_FILE_NAME>
5. a photo of a <embedding>

And there is textual inversion embedding too.

[cansik]

I agree, the quick start guide is really only a quick overview and not really a tutorial or guide on how to prepare your data or train anything in specific (or I could not find it so far). The tool is great, but without further information it's difficult to use.

It seems that training LoRA works for me like the following:

Dataset

1. Prepare your images in a folder
2. Use the dataset tools under the tab tools to prepare the dataset
3. Use the Generate Caption sub-tool to generate captions
4. Set the Initial Caption to the name of the concept (triggerword) (something unique and not already associated in SD)
5. Click Create Captions

Training LoRA

1. Select a LoRA preset from the preset box
2. Under the tab concept add a new concept
3. Click on the concept image and add a name for the concept, datapaths and image augmentations
4. Under the tab sampling, add a sample every 1 epoch with your concept trigger word or prompt
5. Start training and check the tensorboard on how the concept is trained
6. Stop and get the model from models dir

This seemed to work for me, but as @gabeacid already mentioned, I am not sure if this is correct.