controlnet_facefix.py

import torch
from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel
from controlnet_aux import OpenposeDetector, ZoeDetector
from PIL import Image

# ───── Globale Modelle (einmal laden, bleibt im VRAM) ─────
print("Lade OpenPose_faceonly + Depth für perfekte Gesichter/Hände...")

# OpenPose Face-Only Preprocessor + ControlNet
openpose_face = OpenposeDetector.from_pretrained(
    "lllyasviel/ControlNet", model_name="openpose_face"
)

# Depth Preprocessor + ControlNet
depth_processor = ZoeDetector.from_pretrained("lllyasviel/ControlNet")

controlnet_face = ControlNetModel.from_pretrained(
    "lllyasviel/control_v11p_sd15_openpose",
    subfolder="faceonly",
    torch_dtype=torch.float16
).to("cuda")

controlnet_depth = ControlNetModel.from_pretrained(
    "lllyasviel/control_v11f1e_sd15_depth",
    torch_dtype=torch.float16
).to("cuda")

# Pipeline-Cache (wird erst beim ersten Aufruf erstellt)
_facefix_pipe = None

def _get_facefix_pipeline(model_id: str):
    global _facefix_pipe
    if _facefix_pipe is None:
        print(f"Lade Face-Fix-Pipeline mit Modell: {model_id}")
        _facefix_pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained(
            model_id,
            controlnet=[controlnet_face, controlnet_depth],
            torch_dtype=torch.float16,
            safety_checker=None,
            requires_safety_checker=False,
        ).to("cuda")
        _facefix_pipe.enable_xformers_memory_efficient_attention()
        _facefix_pipe.enable_model_cpu_offload()  # spart ~2 GB bei 16 GB Karten
    return _facefix_pipe

def apply_facefix(image: Image.Image, prompt: str, negative_prompt: str, seed: int, base_model_path: str):
    """
    Automatischer 20-Sekunden-Fix für perfekte Gesichter & Hände
    """
    pipe = _get_facefix_pipeline(base_model_path)

    # Control-Images aus dem generierten Bild erzeugen
    pose_img = openpose_face(image)
    depth_img = depth_processor(image)

    fixed_image = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        image=image,
        control_image=[pose_img, depth_img],
        controlnet_conditioning_scale=[0.85, 0.60],   # Face stark, Depth mittel
        strength=0.42,
        num_inference_steps=20,
        guidance_scale=7.0,
        generator=torch.Generator("cuda").manual_seed(seed),
    ).images[0]

    return fixed_image