options/BOOXEL_v0.yaml

model:
  target: BOOXEL.models.BOOXEL_model.BOOXELModel
  params:
    ae_dtype: bf16
    diffusion_dtype: fp16
    scale_factor: 0.13025
    disable_first_stage_autocast: True
    network_wrapper: sgm.modules.diffusionmodules.wrappers.ControlWrapper

    denoiser_config:
      target: sgm.modules.diffusionmodules.denoiser.DiscreteDenoiserWithControl
      params:
        num_idx: 1000
        weighting_config:
          target: sgm.modules.diffusionmodules.denoiser_weighting.EpsWeighting
        scaling_config:
          target: sgm.modules.diffusionmodules.denoiser_scaling.EpsScaling
        discretization_config:
          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization

    control_stage_config:
      target: BOOXEL.modules.BOOXEL_v0.GLVControl
      params:
        adm_in_channels: 2816
        num_classes: sequential
        use_checkpoint: True
        in_channels: 4
        out_channels: 4
        model_channels: 320
        attention_resolutions: [4, 2]
        num_res_blocks: 2
        channel_mult: [1, 2, 4]
        num_head_channels: 64
        use_spatial_transformer: True
        use_linear_in_transformer: True
        transformer_depth: [1, 2, 10]  # 注：第一个未使用（因为 attn_res 从 2 开始）32, 16, 8 --> 64, 32, 16
#        transformer_depth: [1, 1, 4]
        context_dim: 2048
        spatial_transformer_attn_type: softmax-xformers
        legacy: False
        input_upscale: 1

    network_config:
      target: BOOXEL.modules.BOOXEL_v0.LightGLVUNet
      params:
        mode: XL-base
        project_type: ZeroSFT
        project_channel_scale: 2
        adm_in_channels: 2816
        num_classes: sequential
        use_checkpoint: True
        in_channels: 4
        out_channels: 4
        model_channels: 320
        attention_resolutions: [4, 2]
        num_res_blocks: 2
        channel_mult: [1, 2, 4]
        num_head_channels: 64
        use_spatial_transformer: True
        use_linear_in_transformer: True
        transformer_depth: [1, 2, 10]  # 注：第一个未使用（因为 attn_res 从 2 开始）32, 16, 8 --> 64, 32, 16
        context_dim: 2048
        spatial_transformer_attn_type: softmax-xformers
        legacy: False

    conditioner_config:
      target: sgm.modules.GeneralConditionerWithControl
      params:
        emb_models:
          # 交叉连接条件
          - is_trainable: False
            input_key: txt
            target: sgm.modules.encoders.modules.FrozenCLIPEmbedder
            params:
              layer: hidden
              layer_idx: 11
          # 交叉和矢量条件
          - is_trainable: False
            input_key: txt
            target: sgm.modules.encoders.modules.FrozenOpenCLIPEmbedder2
            params:
              arch: ViT-bigG-14
              version: laion2b_s39b_b160k
              freeze: True
              layer: penultimate
              always_return_pooled: True
              legacy: False
          # 向量条件
          - is_trainable: False
            input_key: original_size_as_tuple
            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
            params:
              outdim: 256  # 乘以二
          # 向量条件
          - is_trainable: False
            input_key: crop_coords_top_left
            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
            params:
              outdim: 256  # 乘以二
          # 向量条件
          - is_trainable: False
            input_key: target_size_as_tuple
            target: sgm.modules.encoders.modules.ConcatTimestepEmbedderND
            params:
              outdim: 256  # 乘以二

    first_stage_config:
      target: sgm.models.autoencoder.AutoencoderKLInferenceWrapper
      params:
        ckpt_path: ~
        embed_dim: 4
        monitor: val/rec_loss
        ddconfig:
          attn_type: vanilla-xformers
          double_z: true
          z_channels: 4
          resolution: 256
          in_channels: 3
          out_ch: 3
          ch: 128
          ch_mult: [ 1, 2, 4, 4 ]
          num_res_blocks: 2
          attn_resolutions: [ ]
          dropout: 0.0
        lossconfig:
          target: torch.nn.Identity

    sampler_config:
      target: sgm.modules.diffusionmodules.sampling.RestoreEDMSampler
      params:
        num_steps: 100
        restore_cfg: 4.0
        s_churn: 0
        s_noise: 1.003
        discretization_config:
          target: sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization
        guider_config:
          target: sgm.modules.diffusionmodules.guiders.LinearCFG
          params:
            scale: 7.5
            scale_min: 4.0

    p_p:
        '电影级，高对比度，高度精细，使用哈苏相机拍摄，超精细照片，逼真的最大细节，32K，调色，超高清，极致的细节，皮肤毛孔细节，超清晰度，完美无变形。'
    n_p:
        '绘画，油画，插图，绘图，艺术，素描，动漫，卡通，CG 风格，3D 渲染，虚幻引擎，模糊，混色，不清晰，怪异纹理，丑陋，肮脏，凌乱，质量最差，质量低，框架，水印，签名，JPEG 伪影，变形，低分辨率，过度平滑'

SDXL_CKPT: ckpt_sd_xl_base_1.0/sd_xl_base_1.0_0.9vae.safetensors
BOOXEL_CKPT_F: yanranxiaoxi_booxel/BOOXEL-v0.F.ckpt
BOOXEL_CKPT_Q: yanranxiaoxi_booxel/BOOXEL-v0.Q.ckpt
BOOXEL_CKPT: ~

default_setting:
  s_cfg_Quality: 7.5
  spt_linear_CFG_Quality: 4.0
  s_cfg_Fidelity: 4.0
  spt_linear_CFG_Fidelity: 1.0
  edm_steps: 50