app.py

import gc, random, os
import gradio as gr
import torch, spaces
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler

MODEL_ID = "runwayml/stable-diffusion-v1-5"
DTYPE = torch.float16
_PIPE = None

def get_pipe():
    global _PIPE
    if _PIPE is None:
        # Build on CPU
        pipe = StableDiffusionPipeline.from_pretrained(
            MODEL_ID,
            torch_dtype=DTYPE,
            safety_checker=None,
            use_safetensors=True,
            low_cpu_mem_usage=True,
        )
        pipe.scheduler = DPMSolverMultistepScheduler.from_config(
            pipe.scheduler.config, use_karras_sigmas=True, algorithm_type="dpmsolver++"
        )
        pipe.enable_attention_slicing()
        pipe.enable_vae_slicing()
        pipe.enable_model_cpu_offload()
        _PIPE = pipe
    return _PIPE

def snap8(x: int) -> int:
    x = max(256, min(1024, int(x)))
    return x - (x % 8)

@spaces.GPU(duration=120)
def generate(prompt: str, negative: str, steps: int, cfg: float, width: int, height: int, seed: int):
    pipe = get_pipe()           # stays CPU/offloaded until now
    w, h = snap8(width), snap8(height)

    if int(seed) < 0:
        seed = random.randint(0, 2**31 - 1)
    gen = torch.Generator(device="cuda").manual_seed(int(seed))

    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    gc.collect()

    with torch.autocast(device_type="cuda", dtype=DTYPE):
        out = pipe(
            prompt=str(prompt),
            negative_prompt=str(negative or ""),
            num_inference_steps=int(steps),
            guidance_scale=float(cfg),
            width=w, height=h,
            generator=gen,
        )
    return out.images[0]

with gr.Blocks() as demo:
    gr.Markdown("# ZeroGPU + SD1.5 (minimal)")

    with gr.Tab("Text → Image"):
        prompt  = gr.Textbox(label="Prompt", value="a cozy reading nook, warm sunlight, cinematic lighting, highly detailed")
        negative = gr.Textbox(label="Negative (optional)", value="lowres, blurry, watermark, text")
        steps   = gr.Slider(8, 40, value=28, step=1, label="Steps")
        cfg     = gr.Slider(1.0, 12.0, value=7.0, step=0.5, label="CFG")
        width   = gr.Slider(256, 1024, value=640, step=16, label="Width")
        height  = gr.Slider(256, 1024, value=640, step=16, label="Height")
        seed    = gr.Number(value=-1, precision=0, label="Seed (-1 random)")
        out_img = gr.Image(label="Image", interactive=False)
        gr.Button("Generate").click(generate, [prompt, negative, steps, cfg, width, height, seed], out_img)

if __name__ == "__main__":
    # On Spaces: keep it simple; don’t pass odd kwargs.
    # If you see “localhost is not accessible”, add share=True.
    demo.queue(max_size=12).launch(share=True)