Upload folder using huggingface_hub

convert.py

@@ -0,0 +1,55 @@
+import torch, argparse
+import torch.nn as nn
+from diffusers import AutoencoderKL
+
+def convert_module(model: AutoencoderKL):
+    conv_in = model.encoder.conv_in
+    conv_in_new = nn.Conv2d(
+        4,
+        conv_in.out_channels,
+        conv_in.kernel_size,
+        conv_in.stride,
+        conv_in.padding
+    )
+    with torch.no_grad():
+        conv_in_new.weight[:, :3] = conv_in.weight
+        conv_in_new.weight[:, 3:] = 0       
+        conv_in_new.bias.copy_(conv_in.bias)
+    model.encoder.conv_in = conv_in_new
+
+    conv_out = model.decoder.conv_out
+    conv_out_new = nn.Conv2d(
+        conv_out.in_channels,
+        4,
+        conv_out.kernel_size,
+        conv_out.stride,
+        conv_out.padding
+    )
+    with torch.no_grad():
+        conv_out_new.weight[:3] = conv_out.weight
+        conv_out_new.weight[3:] = 0              
+        conv_out_new.bias[:3] = conv_out.bias
+        conv_out_new.bias[3] = 1
+    model.decoder.conv_out = conv_out_new
+    
+    config = dict(model._internal_dict)
+    config.update({
+        "in_channels": 4,
+        "out_channels": 4,
+    })
+    model._internal_dict = config
+
+    return model
+
+def main():
+    arg_parse = argparse.ArgumentParser()
+    arg_parse.add_argument("--src", type=str, required=True, help="source model path")
+    arg_parse.add_argument("--dst", type=str, required=True, help="destination model path")
+    args = arg_parse.parse_args()
+    
+    vae: AutoencoderKL = AutoencoderKL.from_pretrained(args.src)
+    converted_vae = convert_module(vae)
+    converted_vae.save_pretrained(args.dst)
+
+if __name__ == '__main__':
+    main()

finetune_VAE/config.json

@@ -0,0 +1,37 @@
+{
+  "_class_name": "AutoencoderKL",
+  "_diffusers_version": "0.33.1",
+  "act_fn": "silu",
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": true,
+  "in_channels": 4,
+  "latent_channels": 16,
+  "latents_mean": null,
+  "latents_std": null,
+  "layers_per_block": 2,
+  "mid_block_add_attention": true,
+  "norm_num_groups": 32,
+  "out_channels": 4,
+  "sample_size": 1024,
+  "scaling_factor": 0.3611,
+  "shift_factor": 0.1159,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ],
+  "use_post_quant_conv": false,
+  "use_quant_conv": false
+}

finetune_VAE/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e932691bed6fa7dc679e1b1f9a6c1e27f0d54d2cd79ef25fbfb51743d030e199
+size 167671512

finetune_VAE/finetune_diffusion/pytorch_lora_weights.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:279ffa8385820f7c081e83aba61b75c3cfd9d4d623b5dd232b84d26dfc35a7ff
+size 1434546472