File size: 4,750 Bytes
99a0484 ea5c647 35575bb 99a0484 ea5c647 99a0484 ea5c647 35575bb 99a0484 35575bb 99a0484 35575bb 99a0484 ea5c647 99a0484 ea5c647 3e9c18d 99a0484 ea5c647 3e9c18d 99a0484 ea5c647 35575bb ea5c647 35575bb ea5c647 99a0484 |
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 |
from typing import Optional
import torch
from diffusers import ControlNetModel, StableDiffusionControlNetImg2ImgPipeline
from PIL import Image
import internals.util.image as ImageUtil
from internals.pipelines.commons import AbstractPipeline
from internals.pipelines.controlnets import ControlNet
from internals.pipelines.high_res import HighRes
from internals.pipelines.sdxl_llite_pipeline import SDXLLLiteImg2ImgPipeline
from internals.util import get_generators
from internals.util.config import (
get_base_dimension,
get_hf_cache_dir,
get_is_sdxl,
get_num_return_sequences,
)
class RealtimeDraw(AbstractPipeline):
def load(self, pipeline: AbstractPipeline):
if hasattr(self, "pipe"):
return
if get_is_sdxl():
lite_pipe = SDXLLLiteImg2ImgPipeline()
lite_pipe.load(
pipeline,
[
"https://s3.ap-south-1.amazonaws.com/autodraft.model.assets/models/replicate-xl-llite.safetensors"
],
)
self.pipe = lite_pipe
else:
self.__controlnet_scribble = ControlNetModel.from_pretrained(
"lllyasviel/control_v11p_sd15_scribble",
torch_dtype=torch.float16,
cache_dir=get_hf_cache_dir(),
)
self.__controlnet_seg = ControlNetModel.from_pretrained(
"lllyasviel/control_v11p_sd15_seg",
torch_dtype=torch.float16,
cache_dir=get_hf_cache_dir(),
)
kwargs = {**pipeline.pipe.components} # pyright: ignore
kwargs.pop("image_encoder", None)
self.pipe = StableDiffusionControlNetImg2ImgPipeline(
**kwargs, controlnet=self.__controlnet_seg
).to("cuda")
self.pipe.safety_checker = None
self.pipe2 = StableDiffusionControlNetImg2ImgPipeline(
**kwargs, controlnet=[self.__controlnet_scribble, self.__controlnet_seg]
).to("cuda")
self.pipe2.safety_checker = None
def process_seg(
self,
image: Image.Image,
prompt: str,
negative_prompt: str,
seed: int,
):
if get_is_sdxl():
raise Exception("SDXL is not supported for this method")
generator = get_generators(seed, get_num_return_sequences())
image = ImageUtil.resize_image(image, 512)
img = self.pipe.__call__(
image=image,
control_image=image,
prompt=prompt,
num_inference_steps=15,
negative_prompt=negative_prompt,
generator=generator,
guidance_scale=10,
strength=0.8,
).images[0]
return img
def process_img(
self,
prompt: str,
negative_prompt: str,
seed: int,
image: Optional[Image.Image] = None,
image2: Optional[Image.Image] = None,
):
generator = get_generators(seed, get_num_return_sequences())
b_dimen = get_base_dimension()
if not image:
size = (b_dimen, b_dimen)
if image2:
size = image2.size
image = Image.new("RGB", size, color=0)
if not image2:
size = (b_dimen, b_dimen)
if image:
size = image.size
image2 = Image.new("RGB", size, color=0)
if get_is_sdxl():
size = HighRes.find_closest_sdxl_aspect_ratio(image.size[0], image.size[1])
image = image.resize(size)
torch.manual_seed(seed)
images = self.pipe.__call__(
image=image,
condition_image=image,
negative_prompt=negative_prompt,
prompt=prompt,
seed=seed,
num_inference_steps=10,
width=image.size[0],
height=image.size[1],
)
img = images[0]
else:
image = ImageUtil.resize_image(image, b_dimen)
scribble = ControlNet.scribble_image(image)
image2 = ImageUtil.resize_image(image2, b_dimen)
img = self.pipe2.__call__(
image=image,
control_image=[scribble, image2],
prompt=prompt,
num_inference_steps=15,
negative_prompt=negative_prompt,
guidance_scale=10,
generator=generator,
strength=0.9,
width=image.size[0],
height=image.size[1],
controlnet_conditioning_scale=[1.0, 0.8],
).images[0]
img = ImageUtil.resize_image(img, 512)
return img
|