<h2 id="introduction"><strong>Introduction</strong></h2><p>使用篩選過的 8,000 筆資料集，在 Base 模型下訓練 Full fine-tuning Lora，針對肢體、動作、表情等不同方向準備資料集來調整訓練成果。最後使用分層權重，融合回原有模型。</p><p>原本想融合 8-steps Lora，但最終想保留更多可操作性所以沒有這樣做。所以，可以自由搭配 8-steps 或 4-steps 的加速 Lora 來使用。</p><p>Trained a full fine-tuning LoRA on the Base model using a curated dataset of 8,000 samples, with data specifically structured across different aspects—such as anatomy, poses, and facial expressions—to fine-tune the results. Finally, layer-wise weight merging was applied to fuse it back into the original base model.</p><p>I initially considered merging an 8-step LoRA directly, but ultimately decided against it to preserve greater flexibility and control. Therefore, you are free to pair this model with your preferred 8-step or 4-step acceleration LoRAs.</p><p>厳選した8,000件のデータセットを使用し、Baseモデル上でFull fine-tuning LoRAの学習を行いました。手足・四肢、ポーズ・動作、表情など多角的な観点からデータを準備し、出力結果を調整しています。最終的に、層別重み付け（Layer-wise）を用いて元のベースモデルへマージしました。</p><p>当初は8-steps LoRAのマージも検討しましたが、プロンプト追従性や汎用性（操作の自由度）をより高く維持するため、今回は見送りました。そのため、お好みの8-stepsまたは4-stepsの高速化LoRAと自由に組み合わせてご使用いただけます。</p><p>엄선된 8,000건의 데이터셋을 활용하여 Base 모델 기반으로 Full fine-tuning LoRA 학습을 진행했습니다. 사지 및 신체 표현, 포즈와 동작, 표정 등 다양한 방면으로 데이터셋을 세분화하여 준비함으로써 학습 완성도를 조율했습니다. 최종적으로는 레이어별 가중치 병합(Layer-wise merge)을 적용하여 기존 베이스 모델에 융합했습니다. 원래는 8-steps LoRA를 직접 병합하는 방안도 고려했으나, 사용자의 조작 유연성과 자유도를 최대한 보존하기 위해 적용하지 않았습니다. 따라서 취향에 맞춰 8-steps 또는 4-steps 고속화 가속 LoRA와 자유롭게 조합하여 사용하실 수 있습니다.</p><hr /><h2 id="recommend"><strong>Recommend</strong></h2><p>Resolution: 1.2 - 1.6 Megapixels</p><p>DMD Loras:</p><p><a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd">https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd</a><br /><a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd">https://huggingface.co/lodestones/MOSS-Image-SFT-8-steps-dmd</a></p><hr /><h2 id="merge"><strong>Merge</strong></h2><p>MOSS Image SPF + Base, the layer-wise weights used are as follows:</p><pre><code>video_embedder=0.30
text_embedder.proj=0.30
text_embedder.blocks=0.24
time_embedding=0.36
shared_block_modulation=0.30
layers:
    layers.0=0.25 layers.1=0.25 layers.2=0.25 layers.3=0.25 layers.4=0.25 layers.5=0.25 layers.6=0.25 layers.7=0.25
    layers.8=0.32 layers.9=0.32 layers.10=0.32 layers.11=0.32 layers.12=0.32 layers.13=0.32
    layers.14=0.52 layers.15=0.52 layers.16=0.52 layers.17=0.52 layers.18=0.52 layers.19=0.52 layers.20=0.52 layers.21=0.52
    layers.22=0.64 layers.23=0.64 layers.24=0.64 layers.25=0.64 layers.26=0.64 layers.27=0.64 layers.28=0.64 layers.29=0.64 layers.30=0.64 layers.31=0.64 layers.32=0.64 layers.33=0.64 layers.34=0.64 layers.35=0.64
    layers.36=0.72 layers.37=0.72 layers.38=0.72 layers.39=0.72 layers.40=0.72 layers.41=0.72
final_norm_video.weight=0.30
final_linear_video.weight=0.50
video_head_modulation=0.30</code></pre><p>Full Lora fine-tuning was incorporated, with a batch size of 2, training for 4 epochs on a dataset of 7,780 datasets. The fusion weights were set to 0.7.</p><p>Lora Trainer: <a target="_blank" rel="ugc" href="https://github.com/KohakuBlueleaf/KohakuMossTrainer">https://github.com/KohakuBlueleaf/KohakuMossTrainer</a></p><hr /><h2 id="license"><strong>License</strong></h2><p><strong>Apache-2.0. This is a derivative of Apache-2.0 upstream weights. Keep the upstream notices and state your changes when redistributing.</strong></p>

Files with SHA256: 536a132957c4918a666b005b580d4be68bd76db5b4b600be5485d06da28c2561