app.py

import os, random, re, torch
from typing import List, Tuple
from PIL import Image, ImageDraw, ImageFont
import gradio as gr
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler

# =========================
# SPEED PRESET
# =========================
# Use SD Turbo (1.5) – optimized for very few steps on CPU
DEFAULT_MODEL_ID = "stabilityai/sd-turbo"
MODEL_ID = os.getenv("MODEL_ID", DEFAULT_MODEL_ID)

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16 if torch.cuda.is_available() else torch.float32

# Short NSFW guardrail (block, no blur)
NSFW_TERMS = [
    r"\bnsfw\b", r"\bnude\b", r"\bnudity\b", r"\bsex\b", r"\bexplicit\b", r"\bporn\b",
    r"\bboobs\b", r"\bbutt\b", r"\bass\b", r"\bnaked\b", r"\btits\b",
    r"\b18\+\b", r"\berotic\b", r"\bfetish\b"
]
NSFW_REGEX = re.compile("|".join(NSFW_TERMS), flags=re.IGNORECASE)

def _blocked_tile(reason: str, w=384, h=384) -> Image.Image:
    img = Image.new("RGB", (w, h), (18, 20, 26))
    d = ImageDraw.Draw(img)
    text = f"BLOCKED\n{reason}"
    try:
        font = ImageFont.truetype("DejaVuSans-Bold.ttf", 26)
    except:
        font = ImageFont.load_default()
    box = d.multiline_textbbox((0,0), text, font=font, align="center")
    tw, th = box[2]-box[0], box[3]-box[1]
    d.multiline_text(((w-tw)//2, (h-th)//2), text, font=font, fill=(255,255,255), align="center")
    return img

def _is_nsfw(s: str) -> bool:
    return bool(NSFW_REGEX.search(s or ""))

# -------------------------
# Load pipeline (fast path)
# -------------------------
torch.set_grad_enabled(False)

pipe = StableDiffusionPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=DTYPE,
    safety_checker=None  # let model config handle; we block explicitly on prompts
)

# Turbo still benefits from DPMSolver for CPU
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)

if DEVICE == "cuda":
    pipe = pipe.to("cuda")
    pipe.enable_attention_slicing()
    pipe.enable_vae_slicing()
else:
    pipe = pipe.to("cpu")

# -------------------------
# Generate fn (kept lean)
# -------------------------
def generate(
    prompt: str,
    negative_prompt: str,
    steps: int,
    guidance: float,
    width: int,
    height: int,
    seed: int,
    batch_size: int
) -> Tuple[List[Image.Image], str]:
    if not prompt.strip():
        return [], "Add a prompt first."

    # block obvious NSFW prompts
    if _is_nsfw(prompt) or _is_nsfw(negative_prompt or ""):
        return [_blocked_tile("NSFW prompt detected", width, height)], "Blocked: NSFW prompt."

    # SD-Turbo is designed for tiny step counts + low/zero CFG
    # guard rails on parameters
    steps = max(1, min(int(steps), 12))
    guidance = max(0.0, min(float(guidance), 2.0))

    # Seed
    if seed < 0:
        seed = random.randint(0, 2**31 - 1)
    generator = torch.Generator(device=DEVICE).manual_seed(seed)

    out = pipe(
        prompt=prompt,
        negative_prompt=(negative_prompt or None),
        num_inference_steps=steps,
        guidance_scale=guidance,
        width=width,
        height=height,
        num_images_per_prompt=batch_size,
        generator=generator
    )

    imgs = out.images
    # Some sd-turbo configs may not return nsfw flags; we already block on prompt
    msg = f"Model: {MODEL_ID} • Seed: {seed} • Steps: {steps} • CFG: {guidance} • {width}x{height} • Batch: {batch_size}"
    return imgs, msg

# -------------------------
# UI (defaults tuned for CPU)
# -------------------------
with gr.Blocks(title="VibeForge — Fast (CPU-friendly) Image Gen") as demo:
    gr.Markdown(
        """
# VibeForge ⚒️
**Fast, clean image generation (CPU-friendly).**  
Uses **SD-Turbo** tuned for low steps. NSFW inputs are blocked.
        """
    )

    with gr.Row():
        with gr.Column(scale=3):
            prompt = gr.Textbox(
                label="Prompt",
                placeholder="a neon-lit lighthouse on a stormy cliff at night, cinematic, volumetric fog, high contrast"
            )
            negative = gr.Textbox(label="Negative Prompt", placeholder="low quality, watermark, overexposed")
            with gr.Row():
                steps = gr.Slider(1, 12, value=4, step=1, label="Steps (SD-Turbo sweet spot: 2-6)")
                guidance = gr.Slider(0.0, 2.0, value=0.5, step=0.1, label="CFG (SD-Turbo likes low)")
            with gr.Row():
                width = gr.Dropdown(choices=[384, 448, 512], value=384, label="Width")
                height = gr.Dropdown(choices=[384, 448, 512], value=384, label="Height")
            with gr.Row():
                seed = gr.Number(value=-1, label="Seed (-1 = random)", precision=0)
                batch = gr.Slider(1, 2, value=1, step=1, label="Batch (keep small on CPU)")

            go = gr.Button("Generate", variant="primary")

        with gr.Column(scale=5):
            gallery = gr.Gallery(label="Output", columns=2, height=448)
            info = gr.Markdown()

    go.click(
        fn=generate,
        inputs=[prompt, negative, steps, guidance, width, height, seed, batch],
        outputs=[gallery, info]
    )

if __name__ == "__main__":
    demo.launch()