# app.py
# Prompt Image Editor — Hugging Face Space
# Minimal branding in source so the repo can be published under a subsidiary page


import os
import gradio as gr
from PIL import Image
import torch
from diffusers import StableDiffusionPipeline, StableDiffusionInpaintPipeline
from transformers import logging


logging.set_verbosity_error()


# Environment settings (Spaces: Variables & Secrets)
MODEL_ID = os.getenv("MODEL_ID", "runwayml/stable-diffusion-v1-5")
HF_TOKEN = os.getenv("HF_API_TOKEN") # set as a Secret in your Space if required
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"


def load_pipelines():
print(f"Loading model: {MODEL_ID} on {DEVICE}")
if "inpaint" in MODEL_ID or "img2img" in MODEL_ID:
pipe = StableDiffusionInpaintPipeline.from_pretrained(
MODEL_ID,
revision="fp16",
torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
use_auth_token=HF_TOKEN if HF_TOKEN else None,
)
else:
pipe = StableDiffusionPipeline.from_pretrained(
MODEL_ID,
revision="fp16",
torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
use_auth_token=HF_TOKEN if HF_TOKEN else None,
)
if DEVICE == "cuda":
pipe = pipe.to("cuda")
return pipe


pipe = load_pipelines()


def generate_image(prompt: str, negative_prompt: str, steps: int, guidance: float):
if not prompt:
return None
with torch.autocast("cuda") if DEVICE == "cuda" else torch.no_grad():
out = pipe(prompt=prompt, guidance_scale=guidance, num_inference_steps=steps)
return out.images[0]




def edit_image(init_image, mask, prompt: str, negative_prompt: str, steps: int, guidance: float):
if init_image is None:
return None
if mask is None:
return None
init_img = init_image.convert("RGB")
mask_img = mask.convert("L")
with torch.autocast("cuda") if DEVICE == "cuda" else torch.no_grad():
out = pipe(prompt=prompt, image=init_img, mask_image=mask_img, guidance_scale=guidance, num_inference_steps=steps)
return out.images[0]


with gr.Blocks(title="Prompt Image Editor") as demo:
gr.Markdown("# Prompt Image Editor")
with gr.Row():
with gr.Column(scale=2):
mode = gr.Radio(["Generate", "Edit / Inpaint"], value="Generate", label="Mode")
prompt = gr.Textbox(lines=3, label="Prompt")
negative_prompt = gr.Textbox(lines=2, label="Negative prompt (optional)")
steps = gr.Slider(minimum=10, maximum=60, step=5, value=28, label="Steps")
guidance = gr.Slider(minimum=1.0, maximum=20.0, step=0.5, value=7.5, label="Guidance Scale")
run = gr.Button("Run")
with gr.Column(scale=3):
input_image = gr.Image(type="pil", label="Initial image (for editing)")
mask_image = gr.Image(type="pil", label="Mask (white = edit)")
output = gr.Image(label="Output")


def _run(mode, prompt, negative_prompt, steps, guidance, input_image, mask_image):
try:
if mode == "Generate":
return generate_image(prompt, negative_prompt, steps, guidance)
else:
return edit_image(input_image, mask_image, prompt, negative_prompt, steps, guidance)
except Exception as e:
return Image.new('RGB', (512,512), color=(255,0,0))
demo.launch()