Model's Token after training

[Trauter1]

What's the token of the trained model if I want to use it to generate things?

[
    {
        "instance_prompt":      "1girl Arzaxi",
        "class_prompt":         "1girl",
        "instance_data_dir":    "../../../data/Arzaxi",
        "class_data_dir":       "../../../data/1girl"
    },
]

Is it Instance_prompt? But then if we use filewords, is the token still the same?

[
    {
        "instance_prompt":      "1girl Arzaxi, [filewords]",
        "class_prompt":         "1girl",
        "instance_data_dir":    "../../../data/Arzaxi",
        "class_data_dir":       "../../../data/1girl"
    },
]

And using those settings, is there even a token?
All of the models I trained have heavy amount of bleeding so using token wasn't necessary but I want to eliminate it.
Sorry if im asking stupid question, but I looked online for information about it and can't find definite answer.

[d8ahazard]

The filewords bit is all @sgsdxzy, so I'll defer to them to answer that. ;)

[d8ahazard]

Also, moving to discussions as this is less an issue and more "How does this work?"

[sgsdxzy]

Just think it as you were doing textual inversion, and the addition of [filewords] is just like using textual_inversion_templates/subject_filewords.txt instead of textual_inversion_templates/subject.txt. In both cases, you can invoke your subject just by Arzaxi.

[sgsdxzy]

Just think it as you were doing textual inversion, and the addition of [filewords] is just like using textual_inversion_templates/subject_filewords.txt instead of textual_inversion_templates/subject.txt. In both cases, you can invoke your subject just by Arzaxi.

Thank you for answering! I have one more thing to ask, if we tick this option will our json file look like this? "instance_prompt": "[filewords]" If so, will typing this in instance prompt box Transform it into "instance_prompt": "1girl, [filewords]" ?

No, it's from another pr and totally override [filewords] behavior, your prompt is just ignored.

[Trauter1]

Ok, so I just put 1girl Arzaxi, [filewords] in instance prompt, it should automatically take tags from .txt file with the same name as image, without turning on this option right?

[sgsdxzy]

Ok, so I just put 1girl Arzaxi, [filewords] in instance prompt, it should automatically take tags from .txt file with the same name as image, without turning on this option right?

Yes.

[Trauter1]

Ok, thank you for answering!

[leoftm]

#131
I know this is a multi-post since we had the exact same conversation here, but I would like to hear from someone who knows more about this.
(In filename.txt, each teacher image is set by deepdanbooru to e.g. "1girl, bracelet, earrings, long hair, solo, blue sky, blue background, sky, day, facial mark, necklace, facepaint, black hair, red eyes").

I read about it here.
https://dreambooth.github.io/DreamBooth_files/system.png
https://dreambooth.github.io

If the concept I want to teach is, for example, "Arzaxi", the necessary settings are
The instance prompt is "Arzaxi,[filename.txt]" and
I understood that the required class prompt is "[filename.txt]".
It looks like it is also possible with instance prompt is "Arzaxi ,1girl" and class prommt is "1girl".
Is my understanding of this wrong?

[sgsdxzy]

And those two setting, they are no longer necessary as you can just use [filewords] as your prompt to achieve the same result. If you check them however, they override [filewords] behavior and your prompt is ignored.
They will be removed from ui once #85 is fully merged.

[Trauter1]

@sgsdxzy Started generating class images with those settings :

But looking at them, they don't resemble prompts at all.

Here is example of random prompt from .txt generated by me

Looks like [filewords] is just being used as a token for prompt, instead of being replaced.
db_config.json looks like this:
{"model_name": "Olxc", "scheduler": "ddim", "src": "final-pruned.ckpt [925997e9]", "total_steps": 0, "pretrained_model_name_or_path": "Olxc", "instance_data_dir": "/notebooks/Training datas/AliceEnd", "class_data_dir": "", "instance_prompt": "1girl Olxc, [filewords]", "use_filename_as_label": false, "use_txt_as_label": false, "class_prompt": "1girl, [filewords]", "save_sample_prompt": "", "save_sample_negative_prompt": "", "n_save_sample": 1, "save_guidance_scale": 7.5, "save_infer_steps": 30, "with_prior_preservation": true, "num_class_images": 1000, "output_dir": "/notebooks/stable-diffusion-webui/models/dreambooth/Olxc", "working_dir": "/notebooks/stable-diffusion-webui/models/dreambooth/Olxc/working", "resolution": 512, "center_crop": false, "train_text_encoder": true, "train_batch_size": 1, "sample_batch_size": 1, "num_train_epochs": 1, "max_train_steps": 16000, "gradient_accumulation_steps": 1, "gradient_checkpointing": true, "learning_rate": 1e-06, "scale_lr": false, "lr_scheduler": "constant", "lr_warmup_steps": 0, "use_8bit_adam": true, "adam_beta1": 0.9, "adam_beta2": 0.999, "adam_weight_decay": 0.01, "adam_epsilon": 1e-08, "max_grad_norm": 1, "save_preview_every": 15999, "save_embedding_every": 2000, "mixed_precision": "fp16", "not_cache_latents": true, "concepts_list": "", "use_cpu": false, "pad_tokens": true, "hflip": true, "prior_loss_weight": 1, "seed": null}

Edit: now looking through your older comments. I think you didn't implement [filewords] for class images yet. If so, sorry for pinging you lol

[sgsdxzy]

It should be working. Does it generate the accompanying .txt for your class images?

[Trauter1]

No, going to reinstall the extension, I'm on paperspace notebook so maybe something got broken when updating to latest version.

[Trauter1]

Sadly after updating to latest version, my training won't start no matter what setting I use. I don't even get any error in my cmd.
#142 Seems to have the same problem as me.

[sgsdxzy]

I thinks that's related to b3f8d83 adding some (I think not correct) checks. Need to be fixed by @d8ahazard
It's

if not isset(config.concepts_list) and not config.use_filename_as_label and not config.use_txt_as_label:
        msg = "No instance prompt specified."

I think training should start as long as config.instance_prompt is not empty, not the json file path. It's fixed in my resolving merge conflicts of #85

[Trauter1]

Yeah, using .json fixes the issue. And the class images come with txt file now and look like prompts. My bad for not updating properly