Gorilla Press

此 Flux LyCoris 模型使用 SimpleTuner 在本地进行训练，作为初始训练测试，以探索如何训练姿态。数据集仅包含 26 张图像，主要标签为 "gorilla press"。经过几次尝试，并在不同训练轮次中进行了合并，我发现学习率在 1e-4 到 7e-4 之间效果最佳。虽然该模型并非最稳健或最通用，但大多数情况下能准确还原姿态。

最终训练的配置如下：

{ "--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" }