File size: 8,089 Bytes
e5f91f5 7e328d3 e5f91f5 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 |
try:
import gradio as gr
except Exception:
gr = None
import torch
from PIL import Image
import numpy as np
from PIL import Image
from omegaconf import OmegaConf
import os
import cv2
from diffusers import DDIMScheduler, UniPCMultistepScheduler
from diffusers.models import UNet2DConditionModel
from ref_encoder.latent_controlnet import ControlNetModel
from ref_encoder.adapter import *
from ref_encoder.reference_unet import ref_unet
from utils.pipeline import StableHairPipeline
from utils.pipeline_cn import StableDiffusionControlNetPipeline
from huggingface_hub import hf_hub_download
class StableHair:
def __init__(self, config="./configs/hair_transfer.yaml", device="cuda", weight_dtype=torch.float32) -> None:
print("Initializing Stable Hair Pipeline...")
self.config = OmegaConf.load(config)
self.device = device
# Hugging Face repo with weights
repo_id = "LogicGoInfotechSpaces/new_weights"
# Map config paths to Hugging Face repo structure
# Based on config: pretrained_folder: "./models/stage2"
# encoder_path: "pytorch_model.bin" -> stage2/pytorch_model.bin
# adapter_path: "pytorch_model_1.bin" -> stage2/pytorch_model_1.bin
# controlnet_path: "pytorch_model_2.bin" -> stage2/pytorch_model_2.bin
# bald_converter_path: "./models/stage1/pytorch_model.bin" -> stage1/pytorch_model.bin
# Download weights from Hugging Face
encoder_hf_path = hf_hub_download(repo_id=repo_id, filename="stage2/pytorch_model.bin")
adapter_hf_path = hf_hub_download(repo_id=repo_id, filename="stage2/pytorch_model_1.bin")
controlnet_hf_path = hf_hub_download(repo_id=repo_id, filename="stage2/pytorch_model_2.bin")
bald_converter_hf_path = hf_hub_download(repo_id=repo_id, filename="stage1/pytorch_model.bin")
### Load vae controlnet
unet = UNet2DConditionModel.from_pretrained(self.config.pretrained_model_path, subfolder="unet").to(device)
controlnet = ControlNetModel.from_unet(unet).to(device)
_state_dict = torch.load(controlnet_hf_path, map_location="cpu")
controlnet.load_state_dict(_state_dict, strict=False)
controlnet.to(weight_dtype)
### >>> create pipeline >>> ###
self.pipeline = StableHairPipeline.from_pretrained(
self.config.pretrained_model_path,
controlnet=controlnet,
safety_checker=None,
torch_dtype=weight_dtype,
).to(device)
self.pipeline.scheduler = DDIMScheduler.from_config(self.pipeline.scheduler.config)
### load Hair encoder/adapter
self.hair_encoder = ref_unet.from_pretrained(self.config.pretrained_model_path, subfolder="unet").to(device)
_state_dict = torch.load(encoder_hf_path, map_location="cpu")
self.hair_encoder.load_state_dict(_state_dict, strict=False)
self.hair_adapter = adapter_injection(self.pipeline.unet, device=self.device, dtype=torch.float16, use_resampler=False)
_state_dict = torch.load(adapter_hf_path, map_location="cpu")
self.hair_adapter.load_state_dict(_state_dict, strict=False)
### load bald converter
bald_converter = ControlNetModel.from_unet(unet).to(device)
_state_dict = torch.load(bald_converter_hf_path, map_location="cpu")
bald_converter.load_state_dict(_state_dict, strict=False)
bald_converter.to(dtype=weight_dtype)
del unet
### create pipeline for hair removal
self.remove_hair_pipeline = StableDiffusionControlNetPipeline.from_pretrained(
self.config.pretrained_model_path,
controlnet=bald_converter,
safety_checker=None,
torch_dtype=weight_dtype,
)
self.remove_hair_pipeline.scheduler = UniPCMultistepScheduler.from_config(self.remove_hair_pipeline.scheduler.config)
self.remove_hair_pipeline = self.remove_hair_pipeline.to(device)
### move to fp16
self.hair_encoder.to(weight_dtype)
self.hair_adapter.to(weight_dtype)
print("Initialization Done!")
def Hair_Transfer(self, source_image, reference_image, random_seed, step, guidance_scale, scale, controlnet_conditioning_scale):
prompt = ""
n_prompt = ""
random_seed = int(random_seed)
step = int(step)
guidance_scale = float(guidance_scale)
scale = float(scale)
controlnet_conditioning_scale = float(controlnet_conditioning_scale)
# load imgs
H, W, C = source_image.shape
# generate images
set_scale(self.pipeline.unet, scale)
generator = torch.Generator(device="cuda")
generator.manual_seed(random_seed)
sample = self.pipeline(
prompt,
negative_prompt=n_prompt,
num_inference_steps=step,
guidance_scale=guidance_scale,
width=W,
height=H,
controlnet_condition=source_image,
controlnet_conditioning_scale=controlnet_conditioning_scale,
generator=generator,
reference_encoder=self.hair_encoder,
ref_image=reference_image,
).samples
return sample, source_image, reference_image
def get_bald(self, id_image, scale):
H, W = id_image.size
scale = float(scale)
image = self.remove_hair_pipeline(
prompt="",
negative_prompt="",
num_inference_steps=30,
guidance_scale=1.5,
width=W,
height=H,
image=id_image,
controlnet_conditioning_scale=scale,
generator=None,
).images[0]
return image
model = StableHair(config="./configs/hair_transfer.yaml", weight_dtype=torch.float32)
# Define your ML model or function here
def model_call(id_image, ref_hair, converter_scale, scale, guidance_scale, controlnet_conditioning_scale):
# # Your ML logic goes here
id_image = Image.fromarray(id_image.astype('uint8'), 'RGB')
ref_hair = Image.fromarray(ref_hair.astype('uint8'), 'RGB')
id_image = id_image.resize((512, 512))
ref_hair = ref_hair.resize((512, 512))
id_image_bald = model.get_bald(id_image, converter_scale)
id_image_bald = np.array(id_image_bald)
ref_hair = np.array(ref_hair)
image, source_image, reference_image = model.Hair_Transfer(source_image=id_image_bald,
reference_image=ref_hair,
random_seed=-1,
step=30,
guidance_scale=guidance_scale,
scale=scale,
controlnet_conditioning_scale=controlnet_conditioning_scale
)
image = Image.fromarray((image * 255.).astype(np.uint8))
return id_image_bald, image
# Create a Gradio interface
if gr is not None:
iface = gr.Interface(
fn=model_call,
inputs=[
gr.Image(label="ID Image"),
gr.Image(label="Reference Hair"),
gr.Slider(minimum=0.5, maximum=1.5, value=1, label="Converter Scale"),
gr.Slider(minimum=0.0, maximum=3.0, value=1.0, label="Hair Encoder Scale"),
gr.Slider(minimum=1.1, maximum=3.0, value=1.5, label="CFG"),
gr.Slider(minimum=0.1, maximum=2.0, value=1, label="Latent IdentityNet Scale"),
],
outputs=[
gr.Image(type="pil", label="Bald Result"),
gr.Image(type="pil", label="Transfer Result"),
],
title="Hair Transfer Demo",
description="In general, aligned faces work well, but can also be used on non-aligned faces, and you need to resize to 512 * 512"
)
iface.queue().launch(server_name='0.0.0.0', server_port=7860, share=True) |