TTS / app.py
Jacky2305's picture
use FaceID-PlusV2 SDXL from repo root, deps compatible with gradio 6.32 (hub>=1.16, transformers 5.x)
e2db0d0
Raw History Blame Contribute Delete
3.75 kB
import inspect
import random
import cv2
import numpy as np
import torch
import gradio as gr
import spaces
from insightface.app import FaceAnalysis
from diffusers import StableDiffusionXLPipeline
torch_dtype = torch.float16
# ---- 模块级别加载并搬到 cuda(ZeroGPU 在 @spaces.GPU 外是 CUDA 模拟模式,启动期即可 .to("cuda"))----
pipe = StableDiffusionXLPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch_dtype,
safety_checker=None,
requires_safety_checker=False,
)
# FaceID-PlusV2 (SDXL):上传一张脸的 embedding 即可锁定角色
pipe.load_ip_adapter(
"h94/IP-Adapter-FaceID",
weight_name="ip-adapter-faceid-plusv2_sdxl.bin",
)
pipe.to("cuda")
# ---- 人脸编码器(CPU 推理,避免启动期 CUDA provider 问题)----
face_app = FaceAnalysis(name="antelopev2", root="./", providers=["CPUExecutionProvider"])
face_app.prepare(ctx_id=0, det_size=(640, 640))
MAX_SEED = 2**32 - 1
_CALL_PARAMS = set(inspect.signature(pipe.__call__).parameters)
@spaces.GPU(duration=180)
def generate(face_image, prompt, negative_prompt, seed, steps, guidance, ip_scale):
if face_image is None:
raise gr.Error("请先上传一张人脸参考图(正脸效果最好)")
rgb = np.array(face_image.convert("RGB"))
faces = face_app.get(cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR))
if len(faces) == 0:
raise gr.Error("参考图里检测不到人脸,换一张清晰正脸照试试")
emb = torch.from_numpy(np.asarray(faces[0].normed_embedding, dtype=np.float32))
seed = int(seed)
if seed < 0:
seed = random.randint(0, MAX_SEED)
generator = torch.Generator(device="cuda").manual_seed(seed)
kwargs = dict(
prompt=prompt,
negative_prompt=negative_prompt,
num_inference_steps=int(steps),
guidance_scale=float(guidance),
generator=generator,
)
if "ip_adapter_image_embeds" in _CALL_PARAMS:
kwargs["ip_adapter_image_embeds"] = [emb]
elif "image_embeds" in _CALL_PARAMS:
kwargs["image_embeds"] = emb
if "ip_adapter_scale" in _CALL_PARAMS:
kwargs["ip_adapter_scale"] = float(ip_scale)
image = pipe(**kwargs).images[0]
return image, seed
with gr.Blocks(title="CharGen 固定角色生图") as demo:
gr.Markdown(
"# 🎭 CharGen — 固定角色生图\n"
"上传一张人脸参考图,之后所有生成图都是同一个角色。\n"
"SDXL + IP-Adapter FaceID-PlusV2,ZeroGPU 免费 GPU 额度。"
)
with gr.Row():
with gr.Column():
face = gr.Image(label="参考人脸(必传,正脸最佳)", type="pil")
prompt = gr.Textbox(
label="Prompt",
value="a person smiling, cinematic lighting, 8k, detailed face",
)
neg = gr.Textbox(
label="Negative prompt",
value="(lowres, low quality, worst quality), blurry, deformed, watermark, text, ugly",
)
with gr.Row():
seed = gr.Number(label="Seed(-1 = 随机)", value=-1)
steps = gr.Slider(10, 50, 30, step=1, label="采样步数")
with gr.Row():
guidance = gr.Slider(1, 15, 5, step=0.5, label="CFG")
ip_scale = gr.Slider(0, 1.5, 0.8, step=0.05, label="脸相似度 ip-adapter scale")
btn = gr.Button("生成", variant="primary")
with gr.Column():
out = gr.Image(label="生成结果")
out_seed = gr.Number(label="实际使用的 seed")
btn.click(
generate,
[face, prompt, neg, seed, steps, guidance, ip_scale],
[out, out_seed],
)
demo.launch()