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 FlaxStableDiffusionPipeline

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


pipe, params = FlaxStableDiffusionPipeline.from_pretrained(
    "MuhammadHanif/stable-diffusion-v1-5-high-res", 
    dtype=jnp.bfloat16, 
    use_memory_efficient_attention=True
)

def infer(prompts, negative_prompts, width=1088, height=1088, inference_steps=30, seed=0):
    
    num_samples = 1 #jax.device_count()
    rng = create_key(int(seed))
    rng = jax.random.split(rng, jax.device_count())
    
    prompt_ids = pipe.prepare_inputs([prompts] * num_samples)
    negative_prompt_ids = pipe.prepare_inputs([negative_prompts] * num_samples)
    
    p_params = replicate(params)
    prompt_ids = shard(prompt_ids)
    negative_prompt_ids = shard(negative_prompt_ids)
    
    output = pipe(
        prompt_ids=prompt_ids,
        params=p_params,
        height=height,
        width=width,
        prng_seed=rng,
        num_inference_steps=inference_steps,
        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[0]

prompt_input = gr.inputs.Textbox(
    label="Prompt", 
    placeholder="a highly detailed mansion in the autumn by studio ghibli, makoto shinkai"
)
neg_prompt_input = gr.inputs.Textbox(
    label="Negative Prompt", 
    placeholder=""
)

width_slider = gr.inputs.Slider(
    minimum=512, maximum=2048, default=1088, step=64, label="width"
)

height_slider = gr.inputs.Slider(
    minimum=512, maximum=2048, default=1088, step=64, label="height"
)

inf_steps_input = gr.inputs.Slider(
    minimum=1, maximum=100, default=30, step=1, label="Inference Steps"
)


seed_input = gr.inputs.Number(default=0, label="Seed")

app = gr.Interface(
    fn=infer,
    inputs=[prompt_input, neg_prompt_input, width_slider, height_slider, inf_steps_input, seed_input],
    outputs="image",
    title="Stable Diffusion High Resolution",
    description=(
        "Based on stable diffusion 1.5 and fine-tuned on 576x576 up to 1088x1088 images, "
        "Stable Diffusion High Resolution is compartible with another SD1.5 model and mergeable with other SD1.5 model, "
        "giving other model to generate high resolution images without using upscaler."
    ),
    # examples=[["a highly detailed mansion in the autumn by studio ghibli, makoto shinkai","", 1088, 1088, 30, 0]],
    
)

app.launch()