app/pipeline.py

from diffusers import (
    StableDiffusionControlNetImg2ImgPipeline,
    ControlNetModel,
    LCMScheduler,
    AutoencoderTiny,
)
from compel import Compel
import torch
from utils.canny_gpu import SobelOperator

try:
    import intel_extension_for_pytorch as ipex  # type: ignore
except:
    pass

from pydantic import BaseModel, Field
from PIL import Image
import psutil
import math
import time
import os

from dotenv import load_dotenv
load_dotenv()

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"
default_negative_prompt = "blurry, low quality, render, 3D, oversaturated"

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"

    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(
            640, min=2, max=15, title="Width", disabled=True, hide=True, id="width"
        )
        height: int = Field(
            480, min=2, max=15, title="Height", disabled=True, hide=True, id="height"
        )
        guidance_scale: float = Field(
            1.0,
            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(
            0.2,
            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, device: torch.device, torch_dtype: torch.dtype):
        controlnet_canny = ControlNetModel.from_pretrained(
            controlnet_model, torch_dtype=torch_dtype
        ).to(device)

        self.pipes = {}

        self.pipe = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
            base_model,
            controlnet=controlnet_canny,
        )

        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=False)
        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,
        )

        self.pipe.vae = AutoencoderTiny.from_pretrained(
            taesd_model, torch_dtype=torch_dtype, use_safetensors=True
        ).to(device)

        if bool(os.getenv("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("RGBA", (640, 480))],
                control_image=[Image.new("RGBA", (640, 480))],
        )

    def predict(self, params: "Pipeline.InputParams", image) -> Image.Image:
        generator = torch.manual_seed(params.seed)
        prompt_embeds = self.pipe.compel_proc(params.prompt)
        control_image = self.canny_torch(
            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=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 os.getenv("CONTROL_NET_OVERLAY", True):
            # 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