This Flux LyCoris was trained locally using SimpleTuner as an initial training test to see how to train poses. The dataset is comprised of only 26 images mainly with "gorilla press" being the only tag used. It took a few tries and there were some merges done with various epochs along the way, but I found 1e-4 to 7e-4 learning rates to be the most effective. It's not the most robust or versatile, but it gets the pose most of the time. The training config for the final run: {
 "--resume_from_checkpoint": "latest",
 "--data_backend_config": "config/multidatabackend.json",
 "--aspect_bucket_rounding": 1,
 "--seed": 42,
 "--minimum_image_size": 0,
 "--disable_benchmark": false,
 "--output_dir": "output/models",
 "--lora_type": "lycoris",
 "--lycoris_config": "config/lycoris_config.json",
 "--max_train_steps": 10000,
 "--num_train_epochs": 0,
 "--checkpointing_steps": 250,
 "--checkpoints_total_limit": 20,
 "--model_type": "lora",
 "--pretrained_model_name_or_path": "black-forest-labs/FLUX.1-dev",
 "--model_family": "flux",
 "--train_batch_size": 1,
 "--gradient_checkpointing": "true",
 "--caption_strategy": "textfile",
 "--caption_dropout_probability": 0.0,
 "--resolution_type": "pixel_area",
 "--resolution": 1024,
 "--validation_seed": 42,
 "--validation_steps": 250,
 "--validation_resolution": "1024x1024",
 "--validation_guidance": 3.5,
 "--validation_guidance_rescale": "0.0",
 "--validation_num_inference_steps": "20",
 "--validation_prompt": "gorilla press",
 "--mixed_precision": "bf16",
 "--optimizer": "adamw_bf16",
 "--snr_gamma": "2",
 "--input_perturbation": ".1",
 "--learning_rate": "7e-4",
 "--lr_scheduler": "constant",
 "--lr_warmup_steps": 100,
 "--base_model_precision": "int8-quanto",
 "--validation_torch_compile": "false"
}<pre><code></code></pre>

Gorilla Press

This Flux LyCoris was trained locally using SimpleTuner as an initial training test to see how to train poses. The dataset is comprised of only 26 images mainly with "gorilla press" being the only tag used. It took a few tries and there were some merges done with various epochs along the way, but I found 1e-4 to 7e-4 learning rates to be the most effective. It's not the most robust or versatile, but it gets the pose most of the time.

Files with SHA256: 3da770b5f4fb92dc35f84a80db573fd116c2418f6dfe513554a618e4e9ca6bc4