Can Dreambooth read accompanying text files with descriptions/captions?

[wktra]

Sorry to beat this dead horse, but it seems like this question is like asking if God is real. 😓 No one on several discords can answer this question.

I've tried many settings with no luck. Or maybe it's that I just have no idea if it's working with DB.

Does anyone know the answer?

Can Dreambooth read accompanying text files with descriptions/captions for training images?

[dali99]

yes.

[wktra]

Can't help you I don't use windows

But I use windows and I'm begging you for help.
I'm literally BEGGING you for help.
If you just look and try to guess at what the problem might be, I'll pass it on to ANYONE else who might have this problem.
PLEASE. PLEASE @dali99

[dali99]

It's not like I'm withholding knowledge because you're using windows. I just don't know how your system works...

Cursory googling suggests it's not parsing the path you've provided as a path, why this happens with the string above is something you'll have to reason through. I don't know what kinds of things are permissible or need escaping in the insanity that is windows platform filepaths.

try make it shorter maybe. I'd recommend using simpler filenames and putting your tags in accompanying text files rather than this approach you seem to be using.

The begging isn't really how we do things on these platforms. Post your question, refrain from tagging and notifying people directly. Wait for someone who does know the answer to come along and answer it. Your comments should be mainly adding more information or documenting things you've tried.

Again I can't help you and I can't dedicate my time to holding your hands through troubleshooting. So aiming your question at the general public rather than me, filing a bug report with the error you're getting, is a much more effective way of getting an answer :)

[wktra]

It's not like I'm withholding knowledge because you're using windows. I just don't know how your system works...

Cursory googling suggests it's not parsing the path you've provided as a path, why this happens with the string above is something you'll have to reason through. I don't know what kinds of things are permissible or need escaping in the insanity that is windows platform filepaths.

try make it shorter maybe. I'd recommend using simpler filenames and putting your tags in accompanying text files rather than this approach you seem to be using.

The begging isn't really how we do things on these platforms. Post your question, refrain from tagging and notifying people directly. Wait for someone who does know the answer to come along and answer it. Your comments should be mainly adding more information or documenting things you've tried.

Again I can't help you and I can't dedicate my time to holding your hands through troubleshooting. So aiming your question at the general public rather than me, filing a bug report with the error you're getting, is a much more effective way of getting an answer :)

I appreciate your response, but please know this:

I AM using accompanying text files. It's pulling those tags FROM accompanying text files!

The training images are all in this fashion: 773sloppy (1).png, 773sloppy (1).txt, 773aloppy (2).png, 773sloppy (2).txt, etc etc

[exdownloader]

You might have a file path length issue. I recently made some changes (locally) to the preview saving to avoid a file path bug, where the sanitised prompt would trigger a failure to save previews. To that end, try renaming your files such that the full file path is below 256 characters. So long as there is an accompanying text file with the same file name (excluding extension), it'll read fine, as you've listed.

[wktra]

@exdownloader you've done me a great kindness. Thank you. I'm going to shorten the path see where that leads me. 🙏🤧💖

[cvang187]

Sidenote: It seems you need to enable filepath beyond the default char limit that Windows has by default. https://www.howtogeek.com/266621/how-to-make-windows-10-accept-file-paths-over-260-characters/

This is how I auto caption each instance image and setup the dreambooth extension to use the concepts_list.json file located in sd_dreambooth_extension\dreambooth:

1. use the AutoCaption.ipynb tool from EveryDream-Trainer. Most likely you'll need a local install of Jupyter Notebook or just copy the files to Google Colab as it runs juypter-notebooks. There are other alternatives but I find this one works quite well.

2. Get Windows PowerToy from the Microsoft Store. It has two very useful feature called "PowerRename" and also one called "ImageResizer." ImageResizer allows you to batch/mass resize images to any demension you want thus saving you time using online services or other programs.

3. Use PowerRename after you've auto-captioned each image with EveryDream-Trainer's juypternotebook. PowerRename allows you to select a bunch of files and rename/append/find and replace and use RegularExpression to batch modify any file name of any file type. You need it to add your instance token and class token to the begining of the image file name. In the example below, "opyamato" is the instance token and "woman" is the class token. Following them is the autocaptioned image description from EveryDream-Trainer.

4. If your images are below 512x512 image size, I'd suggest you use the upscalers in Auto1111's web-ui to upscale them by 2x or so and then use the Window's PowerTool - ImageResize to select all your instance image and resize them to 512x512

5. configure your concepts_list.json to something similar to my example below. This is an excerpt from my concept_list.json that is currently setup to train for 11 characters. You can add each { } within the json file for any x amount of character/concept you would like to train dreambooth with. It is important to note, you need to make sure "file_prompt_contents": "Instance + Class + Description", is set for it to read the instance. class, and description from your image file name. I also pasted my full concepts_list.json to paste bin since it doesnt seem like github is playing nicely with the json code snippet below. concepts_list.json

[ { "max_steps": -1, "instance_data_dir": "D:\\onepiece\\op_woman\\nico", "class_data_dir": "D:\\class_image_w", "file_prompt_contents": "Instance + Class + Description", "instance_prompt": "[filewords]", "class_prompt": "[filewords]", "save_sample_prompt": "opnico", "instance_token": "opnico", "class_token": "woman", "num_class_images": 2000, "class_negative_prompt": "", "class_guidance_scale": 7.5, "class_infer_steps": 40, "save_sample_negative_prompt": "", "n_save_sample": 1, "sample_seed": -1, "save_guidance_scale": 7.5, "save_infer_steps": 40 }, { "max_steps": -1, "instance_data_dir": "D:\\onepiece\\op_woman\\yamato", "class_data_dir": "D:\\class_image_w", "file_prompt_contents": "Instance + Class + Description", "instance_prompt": "[filewords]", "class_prompt": "[filewords]", "save_sample_prompt": "opyamato", "instance_token": "opyamato", "class_token": "a woman", "num_class_images": 2000, "class_negative_prompt": "", "class_guidance_scale": 7.5, "class_infer_steps": 40, "save_sample_negative_prompt": "", "n_save_sample": 1, "sample_seed": -1, "save_guidance_scale": 7.5, "save_infer_steps": 40 } ]

[FugueSegue]

TLDR: If you specify "[filewords]" as the string for the class prompt, does it then read text prompt files that correspond to each and every classification image? The Dreambooth extension doesn't generate text prompt files when it automatically generates classification images based on a prompt you can specify.

Text prompt files for each classification image? That's a lot! Sometimes I have over 3,000 classification images for one training.

Up to this point I've been training only one concept per model training. So I haven't bothered using JSON files yet. I've only used the webui to enter parameters for one concept. It automatically creates a JSON file when it begins training.

The default way of generating classification images is to specify a prompt phrase in the Class Prompt text box under the Concepts tab. Then it generates images based on that prompt and puts them in the directory that you specify in the Classification Dataset Directory text box.

I have assumed that if you enter a prompt in the Class Prompt text box that the Dreambooth extension subsequently uses that same prompt with each of the Classification images during training. Do I really have to have a text prompt file for each of the thousands of classification images? If that is the case then it seems terribly redundant. Even if you forego the use of the Concepts tab in favor of a JSON file, you still have to specify a text string for the class_prompt element of the concepts_list, right?

I can imagine that having text prompt files for each classification image might be somehow useful. I think I could manage the automated production of hundreds of classification text prompt files. But doing all of that work would depend on whether or not entering "[filewords]" as the string for the class_prompt element would invoke this course of action.

[d8ahazard]

If you use [filewords] as a string for the class prompt, two things happen.

One - any existing class images with prompts in their filename or an accompanying txt file should have the prompts parsed and used when training.

Two - Any new class images generated will use existing captions from your instance images. The class token and instance token allow us to generate the proper prompt format regardless of what is found where. So, your instance images can be captioned with the subject, or the class, or both, and as long as you set "file prompt contents" properly, instance prompts will have the subject + class, and class prompts will have the subject removed if necessary, and class injected if not there.

If you do not use [filewords] in your class prompt, then whatever is in the prompt box will be used to generate all class images.

[piyarsquare]

In instance prompt, do I enter just [filewords] or xyz style [filewords] if my accompanying text files are Description only? Thank you.

[FollowMarcos]

https://www.reddit.com/r/StableDiffusion/comments/zcr644/make_better_dreambooth_style_models_by_using/

[gitreal2022]

yes, the idea is to help it separate the concepts from each other

I think regs just retrains the concept with the given set of images and it will create a very limited representation of those re-trained original classes. In effect, it will mostly ruin them.