diff --git a/comfy/diffusers_load.py b/comfy/diffusers_load.py index 11d94c34030..a52e0102b73 100644 --- a/comfy/diffusers_load.py +++ b/comfy/diffusers_load.py @@ -1,87 +1,36 @@ import json import os -import yaml -import folder_paths -from comfy.sd import load_checkpoint -import os.path as osp -import re -import torch -from safetensors.torch import load_file, save_file -from . import diffusers_convert +import comfy.sd +def first_file(path, filenames): + for f in filenames: + p = os.path.join(path, f) + if os.path.exists(p): + return p + return None -def load_diffusers(model_path, fp16=True, output_vae=True, output_clip=True, embedding_directory=None): - diffusers_unet_conf = json.load(open(osp.join(model_path, "unet/config.json"))) - diffusers_scheduler_conf = json.load(open(osp.join(model_path, "scheduler/scheduler_config.json"))) +def load_diffusers(model_path, output_vae=True, output_clip=True, embedding_directory=None): + diffusion_model_names = ["diffusion_pytorch_model.fp16.safetensors", "diffusion_pytorch_model.safetensors", "diffusion_pytorch_model.fp16.bin", "diffusion_pytorch_model.bin"] + unet_path = first_file(os.path.join(model_path, "unet"), diffusion_model_names) + vae_path = first_file(os.path.join(model_path, "vae"), diffusion_model_names) - # magic - v2 = diffusers_unet_conf["sample_size"] == 96 - if 'prediction_type' in diffusers_scheduler_conf: - v_pred = diffusers_scheduler_conf['prediction_type'] == 'v_prediction' + text_encoder_model_names = ["model.fp16.safetensors", "model.safetensors", "pytorch_model.fp16.bin", "pytorch_model.bin"] + text_encoder1_path = first_file(os.path.join(model_path, "text_encoder"), text_encoder_model_names) + text_encoder2_path = first_file(os.path.join(model_path, "text_encoder_2"), text_encoder_model_names) - if v2: - if v_pred: - config_path = folder_paths.get_full_path("configs", 'v2-inference-v.yaml') - else: - config_path = folder_paths.get_full_path("configs", 'v2-inference.yaml') - else: - config_path = folder_paths.get_full_path("configs", 'v1-inference.yaml') + text_encoder_paths = [text_encoder1_path] + if text_encoder2_path is not None: + text_encoder_paths.append(text_encoder2_path) - with open(config_path, 'r') as stream: - config = yaml.safe_load(stream) + unet = comfy.sd.load_unet(unet_path) - model_config_params = config['model']['params'] - clip_config = model_config_params['cond_stage_config'] - scale_factor = model_config_params['scale_factor'] - vae_config = model_config_params['first_stage_config'] - vae_config['scale_factor'] = scale_factor - model_config_params["unet_config"]["params"]["use_fp16"] = fp16 + clip = None + if output_clip: + clip = comfy.sd.load_clip(text_encoder_paths, embedding_directory=embedding_directory) - unet_path = osp.join(model_path, "unet", "diffusion_pytorch_model.safetensors") - vae_path = osp.join(model_path, "vae", "diffusion_pytorch_model.safetensors") - text_enc_path = osp.join(model_path, "text_encoder", "model.safetensors") + vae = None + if output_vae: + vae = comfy.sd.VAE(ckpt_path=vae_path) - # Load models from safetensors if it exists, if it doesn't pytorch - if osp.exists(unet_path): - unet_state_dict = load_file(unet_path, device="cpu") - else: - unet_path = osp.join(model_path, "unet", "diffusion_pytorch_model.bin") - unet_state_dict = torch.load(unet_path, map_location="cpu") - - if osp.exists(vae_path): - vae_state_dict = load_file(vae_path, device="cpu") - else: - vae_path = osp.join(model_path, "vae", "diffusion_pytorch_model.bin") - vae_state_dict = torch.load(vae_path, map_location="cpu") - - if osp.exists(text_enc_path): - text_enc_dict = load_file(text_enc_path, device="cpu") - else: - text_enc_path = osp.join(model_path, "text_encoder", "pytorch_model.bin") - text_enc_dict = torch.load(text_enc_path, map_location="cpu") - - # Convert the UNet model - unet_state_dict = diffusers_convert.convert_unet_state_dict(unet_state_dict) - unet_state_dict = {"model.diffusion_model." + k: v for k, v in unet_state_dict.items()} - - # Convert the VAE model - vae_state_dict = diffusers_convert.convert_vae_state_dict(vae_state_dict) - vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()} - - # Easiest way to identify v2.0 model seems to be that the text encoder (OpenCLIP) is deeper - is_v20_model = "text_model.encoder.layers.22.layer_norm2.bias" in text_enc_dict - - if is_v20_model: - # Need to add the tag 'transformer' in advance so we can knock it out from the final layer-norm - text_enc_dict = {"transformer." + k: v for k, v in text_enc_dict.items()} - text_enc_dict = diffusers_convert.convert_text_enc_state_dict_v20(text_enc_dict) - text_enc_dict = {"cond_stage_model.model." + k: v for k, v in text_enc_dict.items()} - else: - text_enc_dict = diffusers_convert.convert_text_enc_state_dict(text_enc_dict) - text_enc_dict = {"cond_stage_model.transformer." + k: v for k, v in text_enc_dict.items()} - - # Put together new checkpoint - sd = {**unet_state_dict, **vae_state_dict, **text_enc_dict} - - return load_checkpoint(embedding_directory=embedding_directory, state_dict=sd, config=config) + return (unet, clip, vae) diff --git a/nodes.py b/nodes.py index 3e4d5240b9c..5e755f14983 100644 --- a/nodes.py +++ b/nodes.py @@ -475,7 +475,7 @@ def load_checkpoint(self, model_path, output_vae=True, output_clip=True): model_path = path break - return comfy.diffusers_load.load_diffusers(model_path, fp16=comfy.model_management.should_use_fp16(), output_vae=output_vae, output_clip=output_clip, embedding_directory=folder_paths.get_folder_paths("embeddings")) + return comfy.diffusers_load.load_diffusers(model_path, output_vae=output_vae, output_clip=output_clip, embedding_directory=folder_paths.get_folder_paths("embeddings")) class unCLIPCheckpointLoader: