The guide is short as you will see: The number of people who do not know that safestensors can be opened as txt to extract the metadata is impressive. They can steal everything, even the number of images used and repeats, the only thing that is inaccessible is the database
Parámetros Importantes:
"ss_sd_model_name": "animefull-final-pruned-fp16.safetensors",
"ss_clip_skip": "2",
"ss_num_train_images": "384",
"ss_tag_frequency": {
"Casas": {
"ss_epoch": "10",
"ss_face_crop_aug_range": "None",
"ss_full_fp16": "False",
"ss_gradient_accumulation_steps": "1",
"ss_gradient_checkpointing": "False",
"ss_learning_rate": "0.0002",
"ss_lowram": "True",
"ss_lr_scheduler": "cosine_with_restarts",
"ss_lr_warmup_steps": "96",
"ss_max_grad_norm": "1.0",
"ss_max_token_length": "225",
"ss_max_train_steps": "1920",
"ss_min_snr_gamma": "5.0",
"ss_mixed_precision": "fp16",
"ss_network_alpha": "16",
"ss_network_dim": "32",
"ss_network_module": "networks.lora",
"ss_new_sd_model_hash": "06bece8e771a9ec16474db4c7d601d2f04fd08d1b1611072e7dd97c18cec3a09",
"ss_noise_offset": "None",
"ss_num_batches_per_epoch": "192",
"ss_cache_latents": "True",
"ss_caption_dropout_every_n_epochs": "0",
"ss_caption_dropout_rate": "0.0",
"ss_caption_tag_dropout_rate": "0.0",
"ss_dataset_dirs": {
"Casas": {
"n_repeats": 8,
"img_count": 48
"ss_num_epochs": "10",
"ss_num_reg_images": "0",
"ss_optimizer": "bitsandbytes.optim.adamw.AdamW8bit",
"ss_output_name": "Casas",
"ss_prior_loss_weight": "1.0",
"ss_sd_model_hash": "0b1893f6",
"ss_sd_scripts_commit_hash": "5050971ac687dca70ba0486a583d283e8ae324e2",
"ss_seed": "42",
"ss_session_id": "1694177070",
"ss_text_encoder_lr": "4e-05",
"ss_training_comment": "None",
"ss_training_finished_at": "1684514891.2831492",
"ss_training_started_at": "1684513546.7678096",
"ss_unet_lr": "0.0002",
"ss_v2": "False",
"sshs_legacy_hash": "d8afa111",
"sshs_model_hash": Hello everyone!
This is Tomas Agilar speaking, and I'm thrilled to have the opportunity to share my work and passion with all of you. If you enjoy what I do and would like to support me, there are a few ways you can do so:
Every donation, whether it's through Ko-fi or my Patreon membership, has a significant impact on my ability to keep creating and sharing content that you love. You are an essential part of my creative journey, and your support fuels me to reach new heights.
From the bottom of my heart, I want to thank you for considering supporting me. Every act of generosity inspires me and gives me the confidence to keep pursuing my dreams and projects. I encourage you to explore the provided links to learn more about my work and find a way to contribute.
Thank you for being here and for being a part of my community. Together, we can make a difference and achieve wonderful things. Have an amazing day!
With gratitude, Tomas Agilar
Description
FAQ
Comments (13)
Stealing????
I didn't know this information was so coveted that inspecting the metadata counts as stealing. Especially in a field where sharing this information benefits the group as a whole. Instead of trying to hoard what the "Best settings" are (plot twist: there are none), why not share this information so that people can make better and higher quality LoRAs for your extremely niche and specific fetishes? Because at the end of the day, that's what it's all about.
Not to mention that all of this info is readily available by clicking the little "info" button when selecting a Lora in the UI. Pretty nifty for finding the trigger word if you forgot.
I think it was meant in jest...
@Polygon Nah, if you read the way it's all written, it really seems like this dude thinks he uncovered some conspiracy and has to blow the whistle about it.
@anon4324 Because there is. Multibillion-dollar corporations will soon be using these models for their own marketing teams. So yes, you and I may only see it as a hobby and fetish, but corporations will want to profit from it. They will hoard the best data for their models, and you or I will not have access to it.
@jasonrat504529 We're talking about information that's just as freely accessible as the prompt information in a PNG. It's not hidden, nor is it obfuscated. It's there for anyone who clicks an icon in the UI. There's nothing for "multibillion-dollar corporations" to steal. They're just slider values and how many times you tagged something.
@jasonrat504529 beat them with cottage creatives, nuff said. If THEY start hiding metadata tho, that will be pain
Is this also useful for seeing the training information on larger safetensors, for training larger models?
We should probably parse this metadata as part of our file processing. Thanks for the heads up.
from safetensors import safe_open
with safe_open('./x.safetensors') as f:
print(f.metadata())
@rockerBOO work in colab?
Here got a web tool to edit/remove all that informations from safetensors file: https://lorainfo.tools/
This is good advice, thanks for this! (BTW, the pic is pretty funny).
Lora that I was training 4 months ago, using civitai's on-site trainer says: ss_sd_model_name":"sd15SD15InpaintingFp16_15.safetensors" , even though I've chosen standard SD1.5 model (that coasts 500 buzz, not a custom for 1000 buzz). No wonder result sucked.