Mont-Saint-Michel, a French monument (PonyXL)版本v1 (ID: 676190)

Mont-Saint-Michel is a French commune in the Manche department of Normandy.
It takes its name from the rocky islet dedicated to Saint Michael, where the Abbey of Mont-Saint-Michel now stands.

I get very good image composition results with Cyberrealistic Pony.
But you can use any other model, as this Dora is trained on PDXL.

Please note that this is my first lora, I'll take any criticism to improve my next creations.

Lora Strength:

A strength of 0.8 is pretty good for a reliable background while leaving room for imagination

If you want something more accurate, you can do [0.8...1] with no problem, but it will be more difficult to incorporate things.

Note for composing an image:

I get better results without adding the dedicated “mntsntmchl” tag. Instead, I use “building” or “castle” plus a few terms depending on where I want to see the character [ shore /bridge /grass /river /water /sky ... ].

maybe I tagged things wrong and my tag absorbed too much

Kohya Settings :

{

  "ss_output_name": "MONT_ST_MICHEL_V1_PDXL",

  "ss_sd_model_name": "ponyDiffusionV6XL_v6StartWithThisOne.safetensors",

  "ss_network_module": "lycoris.kohya",

  "ss_total_batch_size": 6,

  "ss_resolution": "(1024, 1024)",

  "ss_optimizer": "prodigyopt.prodigy.Prodigy",

  "ss_lr_scheduler": "cosine_with_restarts",

  "ss_clip_skip": 2,

  "ss_network_dim": 32,

  "ss_network_alpha": 16,

  "ss_epoch": 6,

  "ss_num_epochs": 7,

  "ss_steps": 420,

  "ss_max_train_steps": 480,

  "ss_learning_rate": 1,

  "ss_text_encoder_lr": 1,

  "ss_unet_lr": 1,

  "ss_shuffle_caption": "False",

  "ss_keep_tokens": 0,

  "ss_flip_aug": "False",

  "ss_noise_offset": "None",

  "ss_adaptive_noise_scale": "None",

  "ss_min_snr_gamma": 5,

  "ss_training_started_at": "2024-07-27T10:57:12.311Z",

  "ss_training_finished_at": "2024-07-27T11:41:03.904Z",

  "training_time": "0h 43m 51s",

  "sshs_model_hash": "16918e1aeddd87400d5ad36010a71562c262ce7ea046742aed858e8f64394fad"

}

描述:

训练词语: mntsntmchl