<h1 id="heading-22">Update 更新日历</h1>📅 2024/1/8:Lion learns densly distributed weights as expected. But we do not consider the performance of the resulting module attractive, so we do not release it. Lion确实按照预期学习到了更稠密的权重，但我们不认为新的模组的性能更具有吸引力，因此我们没有发布该版本。📅 2024/1/8:According to our experience, Lion outperforms AdamW in some cases, an thus we decide to ablate on the optimizer for this module. 根据我们的经验，Lion在某些情况下优于AdamW，因此我们决定对该模块的优化器进行研究。We are re-training this module by Lion instead of AdamW, expecting potentially better performance. 我们正在由Lion替代AdamW重新训练这个模块，期望有更好的表现。<h1 id="heading-2253">Intro. 简介</h1>😨 This is solely a STYLE module for Heaven Burns Red (ヘブンバーンズレッド), not an ALL-IN-ONE (AIO) character package, and you should not expect it to reproduce any character or clothing!😨 这只是Heaven Burns Red (ヘブンバーンズレッド) 的画风模块, 不是一个 ALL-IN-ONE（AIO）角色包，你不应该期望它能复刻任何角色或服装！😋 This model is trained upon ~1700 images for 10 epochs. And thus the author deem that it deserve a ~90MB-parameter size.😋 该模型在大约1700张CG/游戏内资源上训练而成。因此作者认为值得使用90MB大小的参数量来建模。😕 Recommended Inference Settings 推荐的推理设置:<code>CFG: 5.5~7</code><code>Sampler: Euler a</code><code>Steps: 28~35</code><code>High-res.: RealESRGAN_x4plus_anime_6B, 15 steps, 0.55~0.6 denoising strength</code><hr />⚙ Key Training Setting Details 详情核心训练设置:⚠ Due to the large size of the model, the author adopted a relatively conservative hyperparameter setting to promote stability.⚠ 由于模型较大，因此作者采取了比较保守的超参设置来促进稳定性。<pre><code>resolution = "1024"
enable_bucket = true
min_bucket_reso = 256
max_bucket_reso = 2_560
bucket_reso_steps = 64
save_precision = "bf16"
max_train_epochs = 10
train_batch_size = 8
gradient_checkpointing = true
network_train_unet_only = false
network_train_text_encoder_only = false
learning_rate = 0.0001
unet_lr = 0.0001
text_encoder_lr = 0.00001
lr_scheduler = "cosine"
lr_warmup_steps = 500
optimizer_type = "AdamW"
network_module = "lycoris.kohya"
network_dim = 32
network_alpha = 16
network_dropout = 0.1
max_grad_norm = 1
seed = 1_337
clip_skip = 2
mixed_precision = "bf16"
network_args = [ "conv_dim=4", "conv_alpha=2", "dropout=0", "algo=loha" ]</code></pre><hr /><h2 id="heading-2294">Disclaimer for Module</h2>1. Purpose of the Module:The Style LoRA Module ("the Module") provided herein is strictly for educational and research purposes. It is intended to serve as a tool for academic exchange, learning, and for fostering innovation within the field of artificial intelligence and computational creativity.2. No Commercial Use:The user agrees that the Module is not to be used for any commercial purposes. This includes, but is not limited to, selling the Module or any of its components, using the Module or its outputs for commercial gain, or monetizing any aspect of the Module's functionality.3. Copyright Concerns:The user acknowledges that the Module may have been trained on data that could potentially contain copyrighted works. It is the responsibility of the user to ensure that their use of the Module does not infringe upon the rights of any third party or violate any copyright laws.4. No Warranty:The Module is provided 'as is', without any warranty or guarantee of its accuracy, functionality, or performance. The creators of the Module will not be liable for any claims, damages, or other liabilities that may arise from the use of the Module.5. Acknowledgement of Terms:By downloading, accessing, or using the Module, the user acknowledges that they have read this disclaimer, understand its terms, and agree to be bound by them. If you do not agree to these terms, do not download, access, or use the Module.6. Modifications and Amendments:The creators of the Module reserve the right to modify or amend this disclaimer without notice at any time. It is the responsibility of the user to ensure that they are aware of and in compliance with the current terms of this disclaimer as they may be amended or modified.

Files with SHA256: 5c4303d214272de856137a2c5980aeb43d27a00afbe1c4614bf238622eb717c0