doc

README.md

@@ -1,62 +1,33 @@
 # AnimateDiff Model Checkpoints for A1111 SD WebUI
 This repository saves all AnimateDiff models in fp16 & safetensors format for A1111 AnimateDiff users, including
 - motion module (v1-v3)
-- [motion LoRA](#motion-lora) (v2 only, use like any other LoRAs)
-- domain adapter (v3 only, use like any other LoRAs)
-- [sparse ControlNet](#sparse-controlnet) (v3 only, use like any other ControlNets)
+- [motion LoRA](#motion-lora) (v2 only, use like any other LoRA)
+- domain adapter (v3 only, use like any other LoRA)
+- [sparse ControlNet](#sparse-controlnet) (v3 only, use like any other ControlNet)
 
 Unless specified below, you are fine to use models from the [official model repository](https://huggingface.co/guoyww/animatediff/tree/main). I will only convert state dict keys if absolutely necessary.
 
 
 ## Motion LoRA
-Put motion LoRAs to `stable-diffusion-webui/models/Lora` and use motion LoRAs like any other LoRAs you use.
+Put Motion LoRAs to `stable-diffusion-webui/models/Lora` and use Motion LoRAs like any other LoRA you use.
 
-`lora_v2` contains motion LoRAs for AnimateDiff-A1111 v2.0.0. Old motion LoRAs won't work for v2.0.0 and later due to maintenance reason. `lora` will be removed after AnimateDiff-A1111 v2.0.0 is released to master branch.
+[lora_v2](lora_v2) contains motion LoRAs for AnimateDiff-A1111 v2.0.0. I converted state dict keys inside motion LoRAs. Originlal motion LoRAs won't work for AnimateDiff-A1111 v2.0.0 and later due to maintenance reason.
 
-I converted the original state dict via the following code. You may do so if you want to use a motion LoRA from community. Run the following script to make your own motion LoRA checkpoint compatible with AnimateDiff-A1111 v2.0.0 and later.
-```python
-import os, re, torch
-import safetensors.torch
+Use [convert.py](convert.py) in the following way if you want to convert a third-party motion LoRA to be compatible with A1111:
+- Activate your A1111 Python environment first.
+- Command: `python script.py lora [file_path] [save_path]`
+- Replace `[file_path]` with the path to your old LoRA checkpoint.
+- Replace `[save_path]` with the path where you want to save the new LoRA checkpoint.
 
-def convert_mm_name_to_compvis(key):
-    sd_module_key, _, network_part = re.split(r'(_lora\.)', key)
-    sd_module_key = sd_module_key.replace("processor.", "").replace("to_out", "to_out.0")
-    sd_module_key = sd_module_key.replace(".", "_")
-    return f'{sd_module_key}.lora_{network_part}'
 
-file_path = # replace with path to your own old motion LoRA checkpoint
-save_path = # replace with path to your own new motion LoRA checkpoint
-state_dict = safetensors.torch.load_file(file_path) if file_path.endswith(".safetensors") else torch.load(file_path)
-modified_dict = {convert_mm_name_to_compvis(k): v for k, v in state_dict.items()}
-safetensors.torch.save_file(modified_dict, save_path)
-```
+## Sparse ControlNet
+Put Sparse ControlNets to `stable-diffusion-webui/models/ControlNet` and use Sparse ControlNets like any other ControlNet you use.
 
+Like Motion LoRA, I converted state dict keys inside sparse ControlNet. Original sparse ControlNets won't work for A1111 due to maintenance reason.
 
-## Sparse ControlNet
-Put Sparse ControlNets to `stable-diffusion-webui/models/ControlNet` and use Sparse ControlNets like any other ControlNets you use.
-
-Like Motion LoRA, I also converted state dict keys inside sparse ControlNet. Run the following script to make your own sparse ControlNet checkpoint compatible with AnimateDiff-A1111.
-```python
-import torch
-import safetensors.torch
-ad_cn_old = "v3_sd15_sparsectrl_scribble.ckpt" # replace with path to your own old sparse ControlNet checkpoint
-ad_cn_new = "mm_sd15_v3_sparsectrl_scribble.safetensors" # replace with path to your own new sparse ControlNet checkpoint
-normal_cn_path = "diffusion_pytorch_model.fp16.safetensors" # download https://huggingface.co/lllyasviel/control_v11p_sd15_openpose/resolve/main/diffusion_pytorch_model.fp16.safetensors?download=true and replace with the path to this model
-ad_cn = safetensors.torch.load_file(file_path) if file_path.endswith(".safetensors") else torch.load(ad_cn_old)
-normal_cn = safetensors.torch.load_file(normal_cn_path)
-ad_cn_l, ad_cn_m = {}, {}
-for k in ad_cn.keys():
-    if k.startswith("controlnet_cond_embedding"):
-        new_key = k.replace("controlnet_cond_embedding.", "input_hint_block.0.")
-        ad_cn_m[new_key] = ad_cn[k].to(torch.float16)
-    elif not k in normal_cn:
-        if "motion_modules" in k:
-            ad_cn_m[k] = ad_cn[k].to(torch.float16)
-        else:
-            raise Exception(f"{k} not in normal_cn")
-    else:
-        ad_cn_l[k] = ad_cn[k].to(torch.float16)
-ad_cn_l = convert_from_diffuser_state_dict(ad_cn_l)
-ad_cn_l.update(ad_cn_m)
-safetensors.torch.save_file(ad_cn_l, ad_cn_new)
-```
+Use [convert.py](convert.py) in the following way if you want to convert a third-party sparse ControlNet to be compatible with A1111:
+- Activate your A1111 Python environment first.
+- Command: `python script.py controlnet [ad_cn_old] [ad_cn_new] [normal_cn_path]`
+- Replace `[ad_cn_old]` with the path to your old sparse ControlNet checkpoint.
+- Replace `[ad_cn_new]` with the path where you want to save the new sparse ControlNet checkpoint.
+- Replace `[normal_cn_path]` with the path to the normal ControlNet model. Download normal ControlNet from [here](https://huggingface.co/lllyasviel/control_v11p_sd15_openpose/resolve/main/diffusion_pytorch_model.fp16.safetensors?download=true).

convert.py

@@ -0,0 +1,158 @@
+import argparse
+import re
+import torch
+import safetensors.torch
+
+
+def convert_mm_name_to_compvis(key):
+    sd_module_key, _, network_part = re.split(r'(_lora\.)', key)
+    sd_module_key = sd_module_key.replace("processor.", "").replace("to_out", "to_out.0")
+    sd_module_key = sd_module_key.replace(".", "_")
+    return f'{sd_module_key}.lora_{network_part}'
+
+def convert_from_diffuser_state_dict(ad_cn_l):
+    unet_conversion_map = [
+        # (stable-diffusion, HF Diffusers)
+        ("time_embed.0.weight", "time_embedding.linear_1.weight"),
+        ("time_embed.0.bias", "time_embedding.linear_1.bias"),
+        ("time_embed.2.weight", "time_embedding.linear_2.weight"),
+        ("time_embed.2.bias", "time_embedding.linear_2.bias"),
+        ("label_emb.0.0.weight", "add_embedding.linear_1.weight"),
+        ("label_emb.0.0.bias", "add_embedding.linear_1.bias"),
+        ("label_emb.0.2.weight", "add_embedding.linear_2.weight"),
+        ("label_emb.0.2.bias", "add_embedding.linear_2.bias"),
+        ("input_blocks.0.0.weight", "conv_in.weight"),
+        ("input_blocks.0.0.bias", "conv_in.bias"),
+        ("middle_block_out.0.weight", "controlnet_mid_block.weight"),
+        ("middle_block_out.0.bias", "controlnet_mid_block.bias"),
+    ]
+
+    unet_conversion_map_resnet = [
+        # (stable-diffusion, HF Diffusers)
+        ("in_layers.0", "norm1"),
+        ("in_layers.2", "conv1"),
+        ("out_layers.0", "norm2"),
+        ("out_layers.3", "conv2"),
+        ("emb_layers.1", "time_emb_proj"),
+        ("skip_connection", "conv_shortcut"),
+    ]
+
+    unet_conversion_map_layer = []
+    # hardcoded number of downblocks and resnets/attentions...
+    # would need smarter logic for other networks.
+    for i in range(4):
+        # loop over downblocks/upblocks
+
+        for j in range(10):
+            # loop over resnets/attentions for downblocks
+            hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
+            sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
+            unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
+
+            hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
+            sd_down_atn_prefix = f"input_blocks.{3 * i + j + 1}.1."
+            unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
+
+        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
+        sd_downsample_prefix = f"input_blocks.{3 * (i + 1)}.0.op."
+        unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
+
+
+    hf_mid_atn_prefix = "mid_block.attentions.0."
+    sd_mid_atn_prefix = "middle_block.1."
+    unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
+
+    for j in range(2):
+        hf_mid_res_prefix = f"mid_block.resnets.{j}."
+        sd_mid_res_prefix = f"middle_block.{2*j}."
+        unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
+
+    # controlnet specific
+
+    controlnet_cond_embedding_names = ['conv_in'] + [f'blocks.{i}' for i in range(6)] + ['conv_out']
+    for i, hf_prefix in enumerate(controlnet_cond_embedding_names):
+        hf_prefix = f"controlnet_cond_embedding.{hf_prefix}."
+        sd_prefix = f"input_hint_block.{i*2}."
+        unet_conversion_map_layer.append((sd_prefix, hf_prefix))
+
+    for i in range(12):
+        hf_prefix = f"controlnet_down_blocks.{i}."
+        sd_prefix = f"zero_convs.{i}.0."
+        unet_conversion_map_layer.append((sd_prefix, hf_prefix))
+
+
+    def _convert_from_diffuser_state_dict(unet_state_dict):
+        mapping = {k: k for k in unet_state_dict.keys()}
+        for sd_name, hf_name in unet_conversion_map:
+            mapping[hf_name] = sd_name
+        for k, v in mapping.items():
+            if "resnets" in k:
+                for sd_part, hf_part in unet_conversion_map_resnet:
+                    v = v.replace(hf_part, sd_part)
+                mapping[k] = v
+        for k, v in mapping.items():
+            for sd_part, hf_part in unet_conversion_map_layer:
+                v = v.replace(hf_part, sd_part)
+            mapping[k] = v
+        new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items() if k in unet_state_dict}
+        return new_state_dict
+
+    return _convert_from_diffuser_state_dict(ad_cn_l)
+
+
+def lora_conversion(file_path, save_path):
+    state_dict = safetensors.torch.load_file(file_path) if file_path.endswith(".safetensors") else torch.load(file_path)
+    modified_dict = {convert_mm_name_to_compvis(k): v for k, v in state_dict.items()}
+    safetensors.torch.save_file(modified_dict, save_path)
+    print(f"LoRA conversion completed: {save_path}")
+
+
+def controlnet_conversion(ad_cn_old, ad_cn_new, normal_cn_path):
+    ad_cn = safetensors.torch.load_file(ad_cn_old) if ad_cn_old.endswith(".safetensors") else torch.load(ad_cn_old)
+    normal_cn = safetensors.torch.load_file(normal_cn_path)
+    ad_cn_l, ad_cn_m = {}, {}
+    
+    for k in ad_cn.keys():
+        if k.startswith("controlnet_cond_embedding"):
+            new_key = k.replace("controlnet_cond_embedding.", "input_hint_block.0.")
+            ad_cn_m[new_key] = ad_cn[k].to(torch.float16)
+        elif not k in normal_cn:
+            if "motion_modules" in k:
+                ad_cn_m[k] = ad_cn[k].to(torch.float16)
+            else:
+                raise Exception(f"{k} not in normal_cn")
+        else:
+            ad_cn_l[k] = ad_cn[k].to(torch.float16)
+    
+    ad_cn_l = convert_from_diffuser_state_dict(ad_cn_l)
+    ad_cn_l.update(ad_cn_m)
+    safetensors.torch.save_file(ad_cn_l, ad_cn_new)
+    print(f"ControlNet conversion completed: {ad_cn_new}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Script to convert LoRA and ControlNet models.")
+    subparsers = parser.add_subparsers(dest='command')
+
+    # LoRA conversion parser
+    lora_parser = subparsers.add_parser('lora', help='LoRA conversion')
+    lora_parser.add_argument('file_path', type=str, help='Path to the old LoRA checkpoint')
+    lora_parser.add_argument('save_path', type=str, help='Path to save the new LoRA checkpoint')
+
+    # ControlNet conversion parser
+    cn_parser = subparsers.add_parser('controlnet', help='ControlNet conversion')
+    cn_parser.add_argument('ad_cn_old', type=str, help='Path to the old sparse ControlNet checkpoint')
+    cn_parser.add_argument('ad_cn_new', type=str, help='Path to save the new sparse ControlNet checkpoint')
+    cn_parser.add_argument('normal_cn_path', type=str, help='Path to the normal ControlNet model')
+
+    args = parser.parse_args()
+
+    if args.command == 'lora':
+        lora_conversion(args.file_path, args.save_path)
+    elif args.command == 'controlnet':
+        controlnet_conversion(args.ad_cn_old, args.ad_cn_new, args.normal_cn_path)
+    else:
+        parser.print_help()
+
+if __name__ == "__main__":
+    main()