Dynamatrix/DiffBIR-OpenXLab

1target: model.cldm.ControlLDM2params:3  linear_start: 0.000854  linear_end: 0.01205  num_timesteps_cond: 16  log_every_t: 2007  timesteps: 10008  first_stage_key: "jpg"9  cond_stage_key: "txt"10  control_key: "hint"11  image_size: 6412  channels: 413  cond_stage_trainable: false14  conditioning_key: crossattn15  monitor: val/loss_simple_ema16  scale_factor: 0.1821517  use_ema: False18  19  sd_locked: True20  only_mid_control: False21  # Learning rate.22  learning_rate: 1e-423  24  control_stage_config:25    target: model.cldm.ControlNet26    params:27      use_checkpoint: True28      image_size: 32 # unused29      in_channels: 430      hint_channels: 431      model_channels: 32032      attention_resolutions: [ 4, 2, 1 ]33      num_res_blocks: 234      channel_mult: [ 1, 2, 4, 4 ]35      num_head_channels: 64 # need to fix for flash-attn36      use_spatial_transformer: True37      use_linear_in_transformer: True38      transformer_depth: 139      context_dim: 102440      legacy: False41 42  unet_config:43    target: model.cldm.ControlledUnetModel44    params:45      use_checkpoint: True46      image_size: 32 # unused47      in_channels: 448      out_channels: 449      model_channels: 32050      attention_resolutions: [ 4, 2, 1 ]51      num_res_blocks: 252      channel_mult: [ 1, 2, 4, 4 ]53      num_head_channels: 64 # need to fix for flash-attn54      use_spatial_transformer: True55      use_linear_in_transformer: True56      transformer_depth: 157      context_dim: 102458      legacy: False59 60  first_stage_config:61    target: ldm.models.autoencoder.AutoencoderKL62    params:63      embed_dim: 464      monitor: val/rec_loss65      ddconfig:66        #attn_type: "vanilla-xformers"67        double_z: true68        z_channels: 469        resolution: 25670        in_channels: 371        out_ch: 372        ch: 12873        ch_mult:74        - 175        - 276        - 477        - 478        num_res_blocks: 279        attn_resolutions: []80        dropout: 0.081      lossconfig:82        target: torch.nn.Identity83 84  cond_stage_config:85    target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder86    params:87      freeze: True88      layer: "penultimate"89 90  preprocess_config:91    target: model.swinir.SwinIR92    params:93      img_size: 6494      patch_size: 195      in_chans: 396      embed_dim: 18097      depths: [6, 6, 6, 6, 6, 6, 6, 6]98      num_heads: [6, 6, 6, 6, 6, 6, 6, 6]99      window_size: 8100      mlp_ratio: 2101      sf: 8102      img_range: 1.0103      upsampler: "nearest+conv"104      resi_connection: "1conv"105      unshuffle: True106      unshuffle_scale: 8107