app.py

import gradio as gr
import jax
import jax.numpy as jnp
from diffusers import FlaxPNDMScheduler, FlaxStableDiffusionPipeline
from flax.jax_utils import replicate
from flax.training.common_utils import shard

DTYPE = jnp.bfloat16

pipeline, pipeline_params = FlaxStableDiffusionPipeline.from_pretrained(
    "bguisard/stable-diffusion-nano-2-1",
    dtype=DTYPE,
)
if DTYPE != jnp.float32:
    # There is a known issue with schedulers when loading from a pre trained
    # pipeline. We need the schedulers to always use float32.
    # See: https://github.com/huggingface/diffusers/issues/2155
    scheduler, scheduler_params = FlaxPNDMScheduler.from_pretrained(
        pretrained_model_name_or_path="bguisard/stable-diffusion-nano-2-1",
        subfolder="scheduler",
        dtype=jnp.float32,
    )
    pipeline_params["scheduler"] = scheduler_params
    pipeline.scheduler = scheduler


def generate_image(prompt: str, inference_steps: int = 30, prng_seed: int = 0):
    rng = jax.random.PRNGKey(int(prng_seed))
    rng = jax.random.split(rng, jax.device_count())
    p_params = replicate(pipeline_params)

    num_samples = 1
    prompt_ids = pipeline.prepare_inputs([prompt] * num_samples)
    prompt_ids = shard(prompt_ids)

    images = pipeline(
        prompt_ids=prompt_ids,
        params=p_params,
        prng_seed=rng,
        height=128,
        width=128,
        num_inference_steps=int(inference_steps),
        jit=True,
    ).images

    images = images.reshape((num_samples,) + images.shape[-3:])
    images = pipeline.numpy_to_pil(images)
    return images[0]


prompt_input = gr.inputs.Textbox(
    label="Prompt", placeholder="A watercolor painting of a bird"
)
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=generate_image,
    inputs=[prompt_input, inf_steps_input, seed_input],
    outputs="image",
    title="Stable Diffusion Nano",
    description=(
        "Based on stable diffusion and fine-tuned on 128x128 images, "
        "Stable Diffusion Nano allows for fast prototyping of diffusion models, "
        "enabling quick experimentation with easily available hardware."
    ),
    # Some examples were copied from hf.co/spaces/stabilityai/stable-diffusion
    examples=[
        # ["A watercolor painting of a bird", 30, 0],
        [
            "A small cabin on top of a snowy mountain in the style of Disney, artstation",
            25,
            3129302,
        ],
        # ["A mecha robot in a favela in expressionist style", 30, 827198341273],
    ],
)

app.launch()