Spaces:
Running
on
Zero
Running
on
Zero
File size: 4,527 Bytes
55ed985 4811e40 55ed985 4811e40 55ed985 |
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 |
import logging
import torch
from diffusers import (
AutoencoderKL,
EulerDiscreteScheduler,
UNet2DConditionModel,
)
from kolors.models.modeling_chatglm import ChatGLMModel
from kolors.models.tokenization_chatglm import ChatGLMTokenizer
from kolors.models.unet_2d_condition import (
UNet2DConditionModel as UNet2DConditionModelIP,
)
from kolors.pipelines.pipeline_stable_diffusion_xl_chatglm_256 import (
StableDiffusionXLPipeline,
)
from kolors.pipelines.pipeline_stable_diffusion_xl_chatglm_256_ipadapter import ( # noqa
StableDiffusionXLPipeline as StableDiffusionXLPipelineIP,
)
from PIL import Image
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
__all__ = [
"build_text2img_ip_pipeline",
"build_text2img_pipeline",
"text2img_gen",
]
def build_text2img_ip_pipeline(
ckpt_dir: str,
ref_scale: float,
device: str = "cuda",
) -> StableDiffusionXLPipelineIP:
text_encoder = ChatGLMModel.from_pretrained(
f"{ckpt_dir}/text_encoder", torch_dtype=torch.float16
).half()
tokenizer = ChatGLMTokenizer.from_pretrained(f"{ckpt_dir}/text_encoder")
vae = AutoencoderKL.from_pretrained(
f"{ckpt_dir}/vae", revision=None
).half()
scheduler = EulerDiscreteScheduler.from_pretrained(f"{ckpt_dir}/scheduler")
unet = UNet2DConditionModelIP.from_pretrained(
f"{ckpt_dir}/unet", revision=None
).half()
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
f"{ckpt_dir}/../Kolors-IP-Adapter-Plus/image_encoder",
ignore_mismatched_sizes=True,
).to(dtype=torch.float16)
clip_image_processor = CLIPImageProcessor(size=336, crop_size=336)
pipe = StableDiffusionXLPipelineIP(
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
image_encoder=image_encoder,
feature_extractor=clip_image_processor,
force_zeros_for_empty_prompt=False,
)
if hasattr(pipe.unet, "encoder_hid_proj"):
pipe.unet.text_encoder_hid_proj = pipe.unet.encoder_hid_proj
pipe.load_ip_adapter(
f"{ckpt_dir}/../Kolors-IP-Adapter-Plus",
subfolder="",
weight_name=["ip_adapter_plus_general.bin"],
)
pipe.set_ip_adapter_scale([ref_scale])
pipe = pipe.to(device)
# pipe.enable_model_cpu_offload()
# pipe.enable_xformers_memory_efficient_attention()
# pipe.enable_vae_slicing()
return pipe
def build_text2img_pipeline(
ckpt_dir: str,
device: str = "cuda",
) -> StableDiffusionXLPipeline:
text_encoder = ChatGLMModel.from_pretrained(
f"{ckpt_dir}/text_encoder", torch_dtype=torch.float16
).half()
tokenizer = ChatGLMTokenizer.from_pretrained(f"{ckpt_dir}/text_encoder")
vae = AutoencoderKL.from_pretrained(
f"{ckpt_dir}/vae", revision=None
).half()
scheduler = EulerDiscreteScheduler.from_pretrained(f"{ckpt_dir}/scheduler")
unet = UNet2DConditionModel.from_pretrained(
f"{ckpt_dir}/unet", revision=None
).half()
pipe = StableDiffusionXLPipeline(
vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
unet=unet,
scheduler=scheduler,
force_zeros_for_empty_prompt=False,
)
pipe = pipe.to(device)
# pipe.enable_model_cpu_offload()
# pipe.enable_xformers_memory_efficient_attention()
return pipe
def text2img_gen(
prompt: str,
n_sample: int,
guidance_scale: float,
pipeline: StableDiffusionXLPipeline | StableDiffusionXLPipelineIP,
ip_image: Image.Image | str = None,
image_wh: tuple[int, int] = [1024, 1024],
infer_step: int = 50,
ip_image_size: int = 512,
) -> list[Image.Image]:
prompt = "Single " + prompt + ", in the center of the image"
prompt += ", high quality, high resolution, best quality, white background, 3D style," # noqa
logger.info(f"Processing prompt: {prompt}")
kwargs = dict(
prompt=prompt,
height=image_wh[1],
width=image_wh[0],
num_inference_steps=infer_step,
guidance_scale=guidance_scale,
num_images_per_prompt=n_sample,
)
if ip_image is not None:
if isinstance(ip_image, str):
ip_image = Image.open(ip_image)
ip_image = ip_image.resize((ip_image_size, ip_image_size))
kwargs.update(ip_adapter_image=[ip_image])
return pipeline(**kwargs).images
|