文本编码器未训练。
您可以复用基础模型的文本编码器进行推理。

训练设置

• 训练周期: 6
• 训练步数: 3000
• 学习率: 0.0001
  • 学习率调度: 多项式
  • 预热步数: 100
• 最大梯度值: 2.0
• 有效批次大小: 32
  • 微批次大小: 32
  • 梯度累积步数: 1
  • GPU数量: 1
• 梯度检查点: 是
• 预测类型: flow_matching (额外参数=['flow_schedule_auto_shift', 'shift=0.0', 'flow_use_uniform_schedule'])
• 优化器: adamw_bf16
• 可训练参数精度: 纯BF16
• 基础模型精度: 不变
• 标题丢弃概率: 0.1%

LyCORIS配置:

{
    "algo": "lokr",
    "multiplier": 1.0,
    "full_matrix": true,
    "linear_alpha": 1,
    "factor": 16,
    "apply_preset": {
        "target_module": [
            "Attention",
            "FeedForward"
        ],
        "module_algo_map": {
            "Attention": {
                "factor": 16
            },
            "FeedForward": {
                "factor": 8
            }
        }
    }
}

数据集

pseudo-camera-10k-sd3

• 重复次数: 0
• 总图像数: 14102
• 总比例桶数: 1
• 分辨率: 1.048576百万像素
• 裁剪: 是
• 裁剪风格: 居中
• 裁剪比例: 正方形
• 用于正则化数据: 否

推理

import torch
from diffusers import DiffusionPipeline
from lycoris import create_lycoris_from_weights


def download_adapter(repo_id: str):
    import os
    from huggingface_hub import hf_hub_download
    adapter_filename = "pytorch_lora_weights.safetensors"
    cache_dir = os.environ.get('HF_PATH', os.path.expanduser('~/.cache/huggingface/hub/models'))
    cleaned_adapter_path = repo_id.replace("/", "_").replace("\\", "_").replace(":", "_")
    path_to_adapter = os.path.join(cache_dir, cleaned_adapter_path)
    path_to_adapter_file = os.path.join(path_to_adapter, adapter_filename)
    os.makedirs(path_to_adapter, exist_ok=True)
    hf_hub_download(
        repo_id=repo_id, filename=adapter_filename, local_dir=path_to_adapter
    )

    return path_to_adapter_file
    
model_id = 'stabilityai/stable-diffusion-3.5-medium'
adapter_repo_id = 'hmwhwm/simpletuner-lora'
adapter_filename = 'pytorch_lora_weights.safetensors'
adapter_file_path = download_adapter(repo_id=adapter_repo_id)
pipeline = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) # 直接加载为bf16
lora_scale = 1.0
wrapper, _ = create_lycoris_from_weights(lora_scale, adapter_file_path, pipeline.transformer)
wrapper.merge_to()

prompt = "一张猫的照片级真实感图像"
negative_prompt = '模糊、裁剪、丑陋'

## 可选：量化模型以节省显存。
## 注意：训练时未量化模型，因此推理时无需量化。
#from optimum.quanto import quantize, freeze, qint8
#quantize(pipeline.transformer, weights=qint8)
#freeze(pipeline.transformer)
    
pipeline.to('cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu') # 管道已处于目标精度级别
model_output = pipeline(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=20,
    generator=torch.Generator(device='cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu').manual_seed(42),
    width=1024,
    height=1024,
    guidance_scale=4.0,
    skip_guidance_layers=[7, 8, 9],
).images[0]

model_output.save("output.png", format="PNG")