real-time-pix2pix-turbo / pipelines /controlnelSD21Turbo.py
radames's picture
better default for controlnet_scale
d8c251e
raw
history blame
8.49 kB
from diffusers import (
StableDiffusionControlNetImg2ImgPipeline,
ControlNetModel,
LCMScheduler,
AutoencoderTiny,
)
from compel import Compel
import torch
from pipelines.utils.canny_gpu import SobelOperator
try:
import intel_extension_for_pytorch as ipex # type: ignore
except:
pass
import psutil
from config import Args
from pydantic import BaseModel, Field
from PIL import Image
import math
import time
#
taesd_model = "madebyollin/taesd"
controlnet_model = "thibaud/controlnet-sd21-canny-diffusers"
base_model = "stabilityai/sd-turbo"
default_prompt = "Portrait of The Joker halloween costume, face painting, with , glare pose, detailed, intricate, full of colour, cinematic lighting, trending on artstation, 8k, hyperrealistic, focused, extreme details, unreal engine 5 cinematic, masterpiece"
page_content = """
<h1 class="text-3xl font-bold">Real-Time SDv2.1 Turbo</h1>
<h3 class="text-xl font-bold">Image-to-Image ControlNet</h3>
<p class="text-sm">
This demo showcases
<a
href="https://huggingface.co/stabilityai/sd-turbo"
target="_blank"
class="text-blue-500 underline hover:no-underline">SD Turbo</a>
Image to Image pipeline using
<a
href="https://huggingface.co/docs/diffusers/main/en/using-diffusers/sdxl_turbo"
target="_blank"
class="text-blue-500 underline hover:no-underline">Diffusers</a
> with a MJPEG stream server.
</p>
<p class="text-sm text-gray-500">
Change the prompt to generate different images, accepts <a
href="https://github.com/damian0815/compel/blob/main/doc/syntax.md"
target="_blank"
class="text-blue-500 underline hover:no-underline">Compel</a
> syntax.
</p>
"""
class Pipeline:
class Info(BaseModel):
name: str = "controlnet+sd15Turbo"
title: str = "SDv1.5 Turbo + Controlnet"
description: str = "Generates an image from a text prompt"
input_mode: str = "image"
page_content: str = page_content
class InputParams(BaseModel):
prompt: str = Field(
default_prompt,
title="Prompt",
field="textarea",
id="prompt",
)
seed: int = Field(
4402026899276587, min=0, title="Seed", field="seed", hide=True, id="seed"
)
steps: int = Field(
1, min=1, max=15, title="Steps", field="range", hide=True, id="steps"
)
width: int = Field(
512, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
)
height: int = Field(
512, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
)
guidance_scale: float = Field(
1.21,
min=0,
max=10,
step=0.001,
title="Guidance Scale",
field="range",
hide=True,
id="guidance_scale",
)
strength: float = Field(
0.8,
min=0.10,
max=1.0,
step=0.001,
title="Strength",
field="range",
hide=True,
id="strength",
)
controlnet_scale: float = Field(
0.325,
min=0,
max=1.0,
step=0.001,
title="Controlnet Scale",
field="range",
hide=True,
id="controlnet_scale",
)
controlnet_start: float = Field(
0.0,
min=0,
max=1.0,
step=0.001,
title="Controlnet Start",
field="range",
hide=True,
id="controlnet_start",
)
controlnet_end: float = Field(
1.0,
min=0,
max=1.0,
step=0.001,
title="Controlnet End",
field="range",
hide=True,
id="controlnet_end",
)
canny_low_threshold: float = Field(
0.31,
min=0,
max=1.0,
step=0.001,
title="Canny Low Threshold",
field="range",
hide=True,
id="canny_low_threshold",
)
canny_high_threshold: float = Field(
0.125,
min=0,
max=1.0,
step=0.001,
title="Canny High Threshold",
field="range",
hide=True,
id="canny_high_threshold",
)
debug_canny: bool = Field(
False,
title="Debug Canny",
field="checkbox",
hide=True,
id="debug_canny",
)
def __init__(self, args: Args, device: torch.device, torch_dtype: torch.dtype):
controlnet_canny = ControlNetModel.from_pretrained(
controlnet_model, torch_dtype=torch_dtype
).to(device)
self.pipes = {}
if args.safety_checker:
self.pipe = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
base_model,
controlnet=controlnet_canny,
)
else:
self.pipe = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
base_model,
controlnet=controlnet_canny,
safety_checker=None,
)
if args.use_taesd:
self.pipe.vae = AutoencoderTiny.from_pretrained(
taesd_model, torch_dtype=torch_dtype, use_safetensors=True
).to(device)
self.canny_torch = SobelOperator(device=device)
self.pipe.scheduler = LCMScheduler.from_config(self.pipe.scheduler.config)
self.pipe.set_progress_bar_config(disable=True)
self.pipe.to(device=device, dtype=torch_dtype).to(device)
if device.type != "mps":
self.pipe.unet.to(memory_format=torch.channels_last)
if psutil.virtual_memory().total < 64 * 1024**3:
self.pipe.enable_attention_slicing()
self.pipe.compel_proc = Compel(
tokenizer=self.pipe.tokenizer,
text_encoder=self.pipe.text_encoder,
truncate_long_prompts=True,
)
if args.use_taesd:
self.pipe.vae = AutoencoderTiny.from_pretrained(
taesd_model, torch_dtype=torch_dtype, use_safetensors=True
).to(device)
if args.torch_compile:
self.pipe.unet = torch.compile(
self.pipe.unet, mode="reduce-overhead", fullgraph=True
)
self.pipe.vae = torch.compile(
self.pipe.vae, mode="reduce-overhead", fullgraph=True
)
self.pipe(
prompt="warmup",
image=[Image.new("RGB", (768, 768))],
control_image=[Image.new("RGB", (768, 768))],
)
def predict(self, params: "Pipeline.InputParams") -> Image.Image:
generator = torch.manual_seed(params.seed)
prompt_embeds = self.pipe.compel_proc(params.prompt)
control_image = self.canny_torch(
params.image, params.canny_low_threshold, params.canny_high_threshold
)
steps = params.steps
strength = params.strength
if int(steps * strength) < 1:
steps = math.ceil(1 / max(0.10, strength))
last_time = time.time()
results = self.pipe(
image=params.image,
control_image=control_image,
prompt_embeds=prompt_embeds,
generator=generator,
strength=strength,
num_inference_steps=steps,
guidance_scale=params.guidance_scale,
width=params.width,
height=params.height,
output_type="pil",
controlnet_conditioning_scale=params.controlnet_scale,
control_guidance_start=params.controlnet_start,
control_guidance_end=params.controlnet_end,
)
print(f"Time taken: {time.time() - last_time}")
nsfw_content_detected = (
results.nsfw_content_detected[0]
if "nsfw_content_detected" in results
else False
)
if nsfw_content_detected:
return None
result_image = results.images[0]
if params.debug_canny:
# paste control_image on top of result_image
w0, h0 = (200, 200)
control_image = control_image.resize((w0, h0))
w1, h1 = result_image.size
result_image.paste(control_image, (w1 - w0, h1 - h0))
return result_image