| import os |
| import gradio as gr |
| import torch |
| import numpy as np |
| from PIL import Image |
| import cv2 |
|
|
| from diffusers import StableDiffusionControlNetImg2ImgPipeline, ControlNetModel |
| import insightface |
|
|
| |
| device = "cpu" |
| dtype = torch.float32 |
|
|
| BASE_MODEL = "runwayml/stable-diffusion-v1-5" |
| CONTROLNET_MODEL = "lllyasviel/control_v11e_sd15_reference_only" |
|
|
| |
| def pil_to_cv(img): |
| return cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR) |
|
|
| def cv_to_pil(img): |
| return Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) |
|
|
| def make_grid(images, rows, cols): |
| w, h = images[0].size |
| grid = Image.new("RGB", (cols*w, rows*h)) |
| for i, im in enumerate(images): |
| grid.paste(im, box=((i % cols) * w, (i // cols) * h)) |
| return grid |
|
|
| |
| print("🔄 Loading ControlNet...") |
| controlnet = ControlNetModel.from_pretrained(CONTROLNET_MODEL, torch_dtype=dtype) |
|
|
| print("🔄 Loading Stable Diffusion pipeline...") |
| pipe = StableDiffusionControlNetImg2ImgPipeline.from_pretrained( |
| BASE_MODEL, |
| controlnet=controlnet, |
| torch_dtype=dtype, |
| safety_checker=None |
| ).to(device) |
|
|
| print("🔄 Loading InsightFace (antelopev2)...") |
| ins_model = insightface.app.FaceAnalysis(name="antelopev2", providers=["CPUExecutionProvider"]) |
| ins_model.prepare(ctx_id=0, det_size=(640, 640)) |
|
|
| |
| def generate_grid(init_image, prompt, neg_prompt, steps, cfg, strength, seed, width, height): |
| if init_image is None: |
| return None, "❌ Upload foto dulu." |
|
|
| |
| cv_img = pil_to_cv(init_image) |
| faces = ins_model.get(cv_img) |
| if not faces: |
| return None, "❌ Wajah tidak terdeteksi." |
|
|
| |
| faces.sort(key=lambda f: (f.bbox[2]-f.bbox[0])*(f.bbox[3]-f.bbox[1]), reverse=True) |
| x1, y1, x2, y2 = map(int, faces[0].bbox) |
| face_crop = init_image.crop((x1, y1, x2, y2)).resize((width, height)) |
|
|
| |
| g = torch.Generator(device=device) |
| if seed >= 0: |
| g.manual_seed(seed) |
|
|
| images = [] |
| for i in range(9): |
| out = pipe( |
| prompt=prompt, |
| negative_prompt=neg_prompt if neg_prompt.strip() else None, |
| image=init_image.resize((width, height)), |
| control_image=face_crop, |
| num_inference_steps=steps, |
| guidance_scale=cfg, |
| strength=strength, |
| generator=torch.Generator(device=device).manual_seed( |
| seed if seed >= 0 else torch.seed() |
| ) |
| ).images[0] |
| images.append(out) |
|
|
| return make_grid(images, 3, 3), f"✅ Face bbox: {x1},{y1},{x2},{y2}" |
|
|
| |
| with gr.Blocks() as demo: |
| gr.Markdown("## 👤 CPU Face Consistency 3×3 Generator (ControlNet Reference-Only + InsightFace)") |
|
|
| with gr.Row(): |
| with gr.Column(): |
| init_image = gr.Image(label="Upload Foto Wajah", type="pil") |
| prompt = gr.Textbox(label="Prompt", value="studio portrait, soft lighting, ultra detailed skin") |
| neg_prompt = gr.Textbox(label="Negative Prompt", value="blurry, deformed, bad anatomy") |
| steps = gr.Slider(8, 25, 14, step=1, label="Steps") |
| cfg = gr.Slider(3, 12, 7.0, step=0.5, label="CFG Scale") |
| strength = gr.Slider(0.3, 0.8, 0.55, step=0.05, label="Denoising Strength") |
| width = gr.Slider(320, 512, 384, step=32, label="Width") |
| height = gr.Slider(320, 512, 384, step=32, label="Height") |
| seed = gr.Number(value=-1, precision=0, label="Seed (-1 = acak)") |
| run = gr.Button("Generate 3×3") |
| with gr.Column(): |
| out_img = gr.Image(label="Output Grid") |
| log = gr.Markdown() |
|
|
| run.click( |
| fn=generate_grid, |
| inputs=[init_image, prompt, neg_prompt, steps, cfg, strength, seed, width, height], |
| outputs=[out_img, log] |
| ) |
|
|
| if __name__ == "__main__": |
| demo.launch() |
|
|
