app.py

from __future__ import annotations
import math
import random
import spaces
import gradio as gr
import numpy as np
import torch
from PIL import Image
from diffusers import StableDiffusionXLPipeline, EDMEulerScheduler, StableDiffusionXLInstructPix2PixPipeline, AutoencoderKL
from custom_pipeline import CosStableDiffusionXLInstructPix2PixPipeline
from huggingface_hub import hf_hub_download
from huggingface_hub import InferenceClient
from diffusers import StableDiffusion3Pipeline, SD3Transformer2DModel, FlowMatchEulerDiscreteScheduler

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16

repo = "stabilityai/stable-diffusion-3-medium-diffusers"
pipe = StableDiffusion3Pipeline.from_pretrained(repo, torch_dtype=torch.float16).to(device)

help_text = """
To optimize image results:
- Adjust the **Image CFG weight** if the image isn't changing enough or is changing too much. Lower it to allow bigger changes, or raise it to preserve original details.
- Modify the **Text CFG weight** to influence how closely the edit follows text instructions. Increase it to adhere more to the text, or decrease it for subtler changes.
- Experiment with different **random seeds** and **CFG values** for varied outcomes.
- **Rephrase your instructions** for potentially better results.
- **Increase the number of steps** for enhanced edits.
"""

def set_timesteps_patched(self, num_inference_steps: int, device = None):
    self.num_inference_steps = num_inference_steps
    
    ramp = np.linspace(0, 1, self.num_inference_steps)
    sigmas = torch.linspace(math.log(self.config.sigma_min), math.log(self.config.sigma_max), len(ramp)).exp().flip(0)
    
    sigmas = (sigmas).to(dtype=torch.float32, device=device)
    self.timesteps = self.precondition_noise(sigmas)
    
    self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
    self._step_index = None
    self._begin_index = None
    self.sigmas = self.sigmas.to("cpu") 

# Image Editor
edit_file = hf_hub_download(repo_id="stabilityai/cosxl", filename="cosxl_edit.safetensors")
EDMEulerScheduler.set_timesteps = set_timesteps_patched
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
pipe_edit = StableDiffusionXLInstructPix2PixPipeline.from_single_file(
    edit_file, num_in_channels=8, is_cosxl_edit=True, vae=vae, torch_dtype=torch.float16,
)
pipe_edit.scheduler = EDMEulerScheduler(sigma_min=0.002, sigma_max=120.0, sigma_data=1.0, prediction_type="v_prediction")
pipe_edit.to("cuda")

# Generator
@spaces.GPU(duration=30, queue=False)
def king(type ,
        input_image ,
        instruction: str ,
        steps: int = 8,
        randomize_seed: bool = False,
        seed: int = 25,
        text_cfg_scale: float = 7.3,
        image_cfg_scale: float = 1.7,
        width: int = 1024,
        height: int = 1024,
        guidance_scale: float = 6,
        use_resolution_binning: bool = True,
        progress=gr.Progress(track_tqdm=True),
    ):
    if type=="Image Editing" :
        if randomize_seed:
            seed = random.randint(0, 99999)
        text_cfg_scale = text_cfg_scale
        image_cfg_scale = image_cfg_scale
        input_image = input_image

        steps=steps
        generator = torch.manual_seed(seed)
        output_image = pipe_edit(
            instruction, image=input_image,
            guidance_scale=text_cfg_scale, image_guidance_scale=image_cfg_scale,
            num_inference_steps=steps, generator=generator).images[0]
        return seed, output_image
    else :
        if randomize_seed:
            seed = random.randint(0, 99999)
        generator = torch.Generator().manual_seed(seed)
        image = pipe(
            prompt = instruction,
            guidance_scale = 7, 
            num_inference_steps = steps, 
            width = width, 
            height = height,
            generator = generator
        ).images[0]         
        return seed, image

# Prompt classifier
def response(instruction, input_image=None):
    if input_image is None:
        output="Image Generation"
        yield output
    else:
        client = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")
            
        generate_kwargs = dict(
                max_new_tokens=5,
            )
    
        system="[SYSTEM] You will be provided with text, and your task is to classify task is image generation or image editing answer with only task do not say anything else and stop as soon as possible. [TEXT]"
            
        formatted_prompt = system + instruction + "[TASK]"
        stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
        output = ""
        for response in stream:        
            if not response.token.text == "</s>":
                output += response.token.text
        if "editing" in output:
            output = "Image Editing"
        else:
            output = "Image Generation"
        yield output
    return output

css = '''
.gradio-container{max-width: 600px !important}
h1{text-align:center}
footer {
    visibility: hidden
}
'''

examples=[
        [
            "Image Generation",
            None,
            "A Super Car",

        ],
        [
            "Image Editing",
            "./supercar.png",
            "make it red",

        ],
        [
            "Image Editing",
            "./red_car.png",
            "add some snow",

        ],
        [
            "Image Generation",
            None,
            "Kids going o school, Anime style",

        ],
        [
            "Image Generation",
            None,
            "Beautiful Eiffel Tower at Night",

        ],
    ]

with gr.Blocks(css=css) as demo:
    gr.Markdown("# Image Generator Pro")
    with gr.Row():
        with gr.Column(scale=4):
            instruction = gr.Textbox(lines=1, label="Instruction", interactive=True)
        with gr.Column(scale=1):
            type = gr.Dropdown(["Image Generation","Image Editing"], label="Task", value="Image Generation",interactive=True, info="AI will select option based on your query, but if it selects wrong, please choose correct one.")
        with gr.Column(scale=1):
            generate_button = gr.Button("Generate")

    with gr.Row():
        input_image = gr.Image(label="Image", type="pil", interactive=True)

    with gr.Row():
        text_cfg_scale = gr.Number(value=7.3, step=0.1, label="Text CFG", interactive=True)
        image_cfg_scale = gr.Number(value=1.7, step=0.1,label="Image CFG", interactive=True)
        steps = gr.Number(value=25, precision=0, label="Steps", interactive=True)
        randomize_seed = gr.Radio(
                ["Fix Seed", "Randomize Seed"],
                value="Randomize Seed",
                type="index",
                show_label=False,
                interactive=True,
            )
        seed = gr.Number(value=1371, precision=0, label="Seed", interactive=True)

    gr.Examples(
        examples=examples,
        inputs=[type,input_image, instruction],
        fn=king,
        outputs=[input_image],
        cache_examples=False,
    )

    gr.Markdown(help_text)

    instruction.change(fn=response, inputs=[instruction,input_image], outputs=type, queue=False)

    input_image.upload(fn=response, inputs=[instruction,input_image], outputs=type, queue=False)
    
    gr.on(triggers=[
            generate_button.click,
            instruction.submit
        ],
            fn=king,
            inputs=[type,
                input_image,
                instruction,
                steps,
                randomize_seed,
                seed,
                text_cfg_scale,
                image_cfg_scale,
            ],
            outputs=[seed, input_image],
        )

demo.queue(max_size=99999).launch()