app.py

import gradio as gr
import torch
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler, UNet2DConditionModel
from peft import PeftModel

MODEL_ID = "Manojb/stable-diffusion-2-1-base"

LORA_PATHS = {
    "akiec": "./lora/lora_akiec_final",
    "bcc": "./lora/lora_bcc_final",
    "df": "./lora/lora_df_final",
    "mel": "./lora/lora_mel_final",
    "vasc": "./lora/lora_vasc_final",
}

CLASS_NEGATIVE_PROMPTS = {
    "mel": "benign nevus, symmetric, uniform color, regular border, blurry, low quality, artifacts, text, watermark, cartoon, non-dermoscopic, overexposed",
    "bcc": "melanoma, nevus, pigment network, blurry, low quality, artifacts, text, watermark, cartoon, non-dermoscopic, overexposed",
    "akiec": "melanoma, smooth surface, no scale, blurry, low quality, artifacts, text, watermark, cartoon, non-dermoscopic, overexposed",
    "df": "melanoma, irregular border, blue-white veil, blurry, low quality, artifacts, text, watermark, cartoon, non-dermoscopic, overexposed",
    "vasc": "melanoma, pigment network, brown color, blurry, low quality, artifacts, text, watermark, cartoon, non-dermoscopic, overexposed",
}

# Load base pipeline
pipe = StableDiffusionPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float32,
    safety_checker=None,
    requires_safety_checker=False,
)
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe.enable_attention_slicing()

_current_lora = None
_base_unet_state = None

def apply_lora(name):
    global _current_lora, pipe
    if _current_lora == name:
        return
    # Reload fresh UNet then apply LoRA
    pipe.unet = UNet2DConditionModel.from_pretrained(
        MODEL_ID, subfolder="unet", torch_dtype=torch.float32
    )
    pipe.unet = PeftModel.from_pretrained(pipe.unet, LORA_PATHS[name])
    _current_lora = name

def generate(prompt, lora_name, steps):
    apply_lora(lora_name)
    image = pipe(
        prompt,
        negative_prompt=CLASS_NEGATIVE_PROMPTS[lora_name],
        num_inference_steps=int(steps),
        guidance_scale=9.0,
    ).images[0]
    return image

with gr.Blocks() as demo:
    gr.Markdown("# 🧠 DermaDiff — SD 2.1 Generator")
    gr.Markdown("Generate dermoscopy images based on disease class using LoRA")

    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(label="Prompt", value="dermoscopy image of melanoma")
            lora = gr.Dropdown(choices=list(LORA_PATHS.keys()), value="mel", label="Disease Class")
            steps = gr.Slider(10, 50, value=50, step=1, label="Steps")
            btn = gr.Button("Generate", variant="primary")
        with gr.Column():
            output = gr.Image(label="Generated Image")

    btn.click(fn=generate, inputs=[prompt, lora, steps], outputs=output)

demo.launch()