jstep750/animatediff_model

How to make this

• pretrained model: epiCRealism + hyper CFG lora 12steps -> merge with lora weight 0.3
• lora model: AnimateLCM_sd15_t2v_lora.safetensors

import torch

from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler

from animatediff.models.unet import UNet3DConditionModel
from animatediff.models.sparse_controlnet import SparseControlNetModel
from animatediff.pipelines.pipeline_animation import AnimationPipeline
from animatediff.utils.util import load_weights

sdpipe = StableDiffusionPipeline.from_single_file(pretrained_model_path, use_safetensors=True, add_watermarker=False).to(dtype=torch.float16)
sdpipe.load_lora_weights(lora_model_path)
sdpipe.fuse_lora(lora_scale=0.3)

text_encoder = sdpipe.text_encoder.cuda()
vae          = sdpipe.vae.cuda()
tokenizer    = sdpipe.tokenizer

unet_additional_kwargs = params["unet_additional_kwargs"]
controlnet_additional_kwargs = params["controlnet_additional_kwargs"]

unet = UNet3DConditionModel.from_pretrained_2d(pretrained_model_path, subfolder="unet", unet_config=sdpipe.unet.config, unet_additional_kwargs=unet_additional_kwargs).cuda()
unet.config.num_attention_heads = 8
unet.config.projection_class_embeddings_input_dim = None
unet.to(dtype=torch.float16)

controlnet = SparseControlNetModel.from_unet(unet, controlnet_additional_kwargs=controlnet_additional_kwargs)
controlnet_path = "models/motion_module/v3_sd15_sparsectrl_rgb.ckpt"

print(f"loading controlnet checkpoint from {controlnet_path} ...")
controlnet_state_dict = torch.load(controlnet_path, map_location="cpu")
controlnet_state_dict = controlnet_state_dict["controlnet"] if "controlnet" in controlnet_state_dict else controlnet_state_dict
controlnet_state_dict = {name: param for name, param in controlnet_state_dict.items() if "pos_encoder.pe" not in name}
controlnet_state_dict.pop("animatediff_config", "")
controlnet.load_state_dict(controlnet_state_dict)
controlnet.to(dtype=torch.float16)
controlnet.cuda()

pipe = load_weights(
    pipeline,
    # motion module
    motion_module_path         = "models/Motion_Module/v3_sd15_mm.ckpt", 
    motion_module_lora_configs = [],
    # domain adapter
    adapter_lora_path          = "models/Motion_Module/v3_sd15_adapter.ckpt",
    adapter_lora_scale         = 1.0,
    # image layers
    dreambooth_model_path      = pretrained_model_path, 
    lora_model_path            = "",
    lora_alpha                 = 0.8,
).to("cuda")
pipe.to(dtype=torch.float16)
pipe.enable_vae_slicing()
pipe.enable_model_cpu_offload()

pipe.scheduler = DPMSolverMultistepScheduler(
    beta_start = 0.00075,
    beta_end = 0.0145,
    beta_schedule = "linear",
    use_karras_sigmas = True,
)

How to use this

controlnet_additional_kwargs = params["controlnet_additional_kwargs"]

pipe = AnimationPipeline.from_pretrained("models/animatediff_model")
unet = pipe.unet
vae = pipe.vae

unet.config.num_attention_heads = 8
unet.config.projection_class_embeddings_input_dim = None
unet.to(dtype=torch.float16)

controlnet = SparseControlNetModel.from_unet(unet, controlnet_additional_kwargs=controlnet_additional_kwargs)
controlnet_path = "./models/motion_module/v3_sd15_sparsectrl_rgb.ckpt"

print(f"loading controlnet checkpoint from {controlnet_path} ...")
controlnet_state_dict = torch.load(controlnet_path, map_location="cpu")
controlnet_state_dict = controlnet_state_dict["controlnet"] if "controlnet" in controlnet_state_dict else controlnet_state_dict
controlnet_state_dict = {name: param for name, param in controlnet_state_dict.items() if "pos_encoder.pe" not in name}
controlnet_state_dict.pop("animatediff_config", "")
controlnet.load_state_dict(controlnet_state_dict)
controlnet.to(dtype=torch.float16)
controlnet.cuda()

pipe.controlnet = controlnet

without_xformers = False
if is_xformers_available() and (not without_xformers):
    unet.enable_xformers_memory_efficient_attention()
    if controlnet is not None: 
        print("\nenable_xformers_memory_efficient_attention\n")
        controlnet.enable_xformers_memory_efficient_attention()

pipe.to(dtype=torch.float16)
pipe.enable_vae_slicing()
pipe.enable_model_cpu_offload()
pipe.to("cuda")