app.py

import gradio as gr
import jax
import numpy as np
import jax.numpy as jnp
from flax.jax_utils import replicate
from flax.training.common_utils import shard
from PIL import Image
from diffusers import FlaxStableDiffusionControlNetPipeline, FlaxControlNetModel
import cv2

def create_key(seed=0):
    return jax.random.PRNGKey(seed)

def crop_and_resize(pilimg, size=512):
    """
    Will downsample or upsample as necessary.
    """
    width, height = pilimg.size
    minsize = min(width, height)
    x0 = (width - height) // 2 if width > height else 0
    y0 = (height - width) // 2 if height > width else 0
    pilimg = pilimg.crop((x0, y0, x0 + minsize, y0 + minsize))
    pilimg = pilimg.resize((size, size), resample=Image.LANCZOS)
    return pilimg

def canny_filter(image):
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    blurred_image = cv2.GaussianBlur(gray_image, (5, 5), 0)
    edges_image = cv2.Canny(blurred_image, 50, 150)
    return edges_image

# load control net and stable diffusion v1-5
controlnet, controlnet_params = FlaxControlNetModel.from_pretrained(
    "jax-diffusers-event/canny-coyo1m", dtype=jnp.bfloat16
)
pipe, params = FlaxStableDiffusionControlNetPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5", controlnet=controlnet, revision="flax", dtype=jnp.bfloat16
)

def infer(prompts, negative_prompts, image):
    params["controlnet"] = controlnet_params

    # image is a numpy array, we'll convert to PIL to resize and back to numpy
    image = Image.fromarray(image)
    image = crop_and_resize(image)
    image = np.array(image)
    
    num_samples = 1 #jax.device_count()
    rng = create_key(0)
    rng = jax.random.split(rng, jax.device_count())
    im = canny_filter(image)
    canny_image = Image.fromarray(im)
    
    prompt_ids = pipe.prepare_text_inputs([prompts] * num_samples)
    negative_prompt_ids = pipe.prepare_text_inputs([negative_prompts] * num_samples)
    processed_image = pipe.prepare_image_inputs([canny_image] * num_samples)
    
    p_params = replicate(params)
    prompt_ids = shard(prompt_ids)
    negative_prompt_ids = shard(negative_prompt_ids)
    processed_image = shard(processed_image)
    
    output = pipe(
        prompt_ids=prompt_ids,
        image=processed_image,
        params=p_params,
        prng_seed=rng,
        num_inference_steps=50,
        neg_prompt_ids=negative_prompt_ids,
        jit=True,
    ).images
    
    output_images = pipe.numpy_to_pil(np.asarray(output.reshape((num_samples,) + output.shape[-3:])))
    return output_images

gr.Interface(infer, inputs=["text", "text", "image"], outputs="gallery").launch()