dingvero/hfut-bishe
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update

Browse Source
This commit is contained in:
DingVero
2026-05-14 18:37:18 +08:00
parent 5e6d8046e1
commit 7a2ffc91a2
15 changed files with 200 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files
client/gui/dialogs/BlackholeResolveDialog.cpp
+1 -10
View File
@@ -25,14 +25,6 @@ QPushButton* makeAlgoButton(const QString& title, QWidget* parent) {
} // namespace
QString BlackholeResolveDialog::defaultModelInpaintPrompt() {
return QStringLiteral(
"中国画局部，需去除人物并只补背景：请严格依据掩膜外已见的笔墨皴擦、设色层次、留白与纸绢肌理，"
"在掩膜内自然延展，补全为山石、坡岸、云水、苔点或空灵留白等，与外缘笔法、墨色与气韵连贯衔接；"
"禁止出现人物轮廓、肢体、衣褶、五官、手足、发丝、肤色、投影及去人后的模糊晕带、鬼影边缘与色块台阶，"
"整体须像原画从未画入人物一样自然，不留修补感。");
}
BlackholeResolveDialog::BlackholeResolveDialog(const QString& blackholeName, QWidget* parent)
: QDialog(parent),
m_blackholeName(blackholeName) {
@@ -142,8 +134,7 @@ void BlackholeResolveDialog::buildDetailPage() {
pLay->setSpacing(8);
m_promptEdit = new QPlainTextEdit(panel);
m_promptEdit->setPlainText(BlackholeResolveDialog::defaultModelInpaintPrompt());
m_promptEdit->setPlaceholderText(QStringLiteral("可按画面增删；若全部删空，请求时仍使用内置默认描述"));
m_promptEdit->setPlaceholderText(QStringLiteral("可选"));
m_promptEdit->setMinimumHeight(72);
pLay->addWidget(m_promptEdit);
client/gui/dialogs/BlackholeResolveDialog.h
-3
View File
@@ -19,9 +19,6 @@ public:
Algorithm selectedAlgorithm() const { return m_selectedAlgorithm; }
QString promptText() const;
/// 模型补全（LaMa 等）内置正向提示：中国画去人物、补背景纹理，无人物残留。
[[nodiscard]] static QString defaultModelInpaintPrompt();
private:
void buildSelectPage();
void buildDetailPage();
client/gui/main_window/MainWindow.cpp
+4 -12
View File
@@ -5050,21 +5050,13 @@ void MainWindow::showBlackholeContextMenu(const QPoint& globalPos, const QString
client->setBaseUrl(QUrl(base));
QString immediateErr;
QString prompt = dlg.promptText();
if (prompt.isEmpty()) {
prompt = BlackholeResolveDialog::defaultModelInpaintPrompt();
}
const QString prompt = dlg.promptText();
QNetworkReply* reply = client->inpaintAsync(
cropPng,
maskPng,
// LaMa（物体移除 / 擦除）
QStringLiteral("lama"),
QStringLiteral("flux_fill"),
prompt,
// 抑制人物与去人常见伪影（中英混合便于通用后端）
QStringLiteral(
"person, human, face, body, hands, feet, portrait, silhouette, skin, clothing folds, hair, "
"shadow, ghosting, blur halo, seam, banding, watermark, text, logo, signature, jpeg artifacts"),
// 去人物场景：strength 太低容易“保留人物残影”；适当提高更干净
QString(),
0.72,
1024,
&immediateErr);
@@ -5077,7 +5069,7 @@ void MainWindow::showBlackholeContextMenu(const QPoint& globalPos, const QString
}
auto* task = new CancelableTaskDialog(QStringLiteral("黑洞修复"),
QStringLiteral("正在请求后端模型补全，请稍候……"),
QStringLiteral("补全中…"),
this);
task->setAttribute(Qt::WA_DeleteOnClose, true);
connect(task, &CancelableTaskDialog::canceled, this, [reply, task]() {
python_server/.env
+1
View File
@@ -0,0 +1 @@
HF_TOKEN=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
python_server/app/routes/inpaint.py
+7 -3
View File
@@ -15,7 +15,7 @@ router = APIRouter(tags=["inpaint"])
@router.post("/inpaint")
def inpaint(req: InpaintRequest) -> Dict[str, Any]:
try:
model_name = req.model_name or "sdxl_inpaint"
model_name = req.model_name or "flux_fill"
pil = b64_to_pil_image(req.image_b64).convert("RGB")
if req.mask_b64:
@@ -24,13 +24,17 @@ def inpaint(req: InpaintRequest) -> Dict[str, Any]:
mask = default_half_mask(pil)
predictor = get_inpaint_predictor(model_name)
call_kw: Dict[str, Any] = {"strength": req.strength, "max_side": req.max_side}
if req.guidance_scale is not None:
call_kw["guidance_scale"] = req.guidance_scale
if req.num_inference_steps is not None:
call_kw["num_inference_steps"] = req.num_inference_steps
out = predictor(
pil,
mask,
req.prompt or "",
req.negative_prompt or "",
strength=req.strength,
max_side=req.max_side,
**call_kw,
)
out_dir = OUTPUT_DIR / "inpaint"
python_server/app/routes/meta.py
+11 -11
View File
@@ -43,19 +43,19 @@ def get_models() -> Dict[str, Any]:
"mask2former": {"name": "Mask2Former (not implemented)"},
},
"inpaint": {
"copy": {"name": "Copy (no-op)", "params": []},
"sdxl_inpaint": {
"name": "SDXL Inpaint",
"params": [
{"id": "prompt", "label": "提示词", "optional": True},
],
"flux_fill": {
"name": "FLUX.1 Fill (default)",
"params": [{"id": "prompt", "label": "prompt", "optional": True}],
},
"lama": {"name": "LaMa (Erase / Remove Object)", "params": []},
"sdxl_inpaint": {
"name": "sdxl_inpaint",
"params": [{"id": "prompt", "label": "prompt", "optional": True}],
},
"lama": {"name": "lama", "params": []},
"copy": {"name": "copy (debug)", "params": []},
"controlnet": {
"name": "ControlNet Inpaint (canny)",
"params": [
{"id": "prompt", "label": "提示词", "optional": True},
],
"name": "controlnet",
"params": [{"id": "prompt", "label": "prompt", "optional": True}],
},
},
"animation": {
python_server/app/schemas.py
+2
View File
@@ -52,6 +52,8 @@ class InpaintRequest(ImageInput):
negative_prompt: Optional[str] = Field("", description="负向 prompt")
mask_b64: Optional[str] = Field(None, description="mask PNG base64（可选）")
max_side: int = Field(1024, ge=128, le=2048)
guidance_scale: Optional[float] = Field(None, ge=0.0, le=200.0, description="扩散 CFG；不传则各后端默认")
num_inference_steps: Optional[int] = Field(None, ge=1, le=150, description="采样步数；不传则各后端默认")
class AnimateRequest(BaseModel):
python_server/app/services/predictors.py
+5
View File
@@ -76,6 +76,11 @@ def get_inpaint_predictor(model_name: str):
_inpaint_cache[model_name] = _copy
return _copy
if model_name == "flux_fill":
pred, _ = build_inpaint_predictor(UnifiedInpaintConfig(backend=InpaintBackend.FLUX_FILL))
_inpaint_cache[model_name] = pred
return pred
if model_name == "sdxl_inpaint":
pred, _ = build_inpaint_predictor(UnifiedInpaintConfig(backend=InpaintBackend.SDXL_INPAINT))
_inpaint_cache[model_name] = pred
python_server/config.py
+16 -7
View File
@@ -86,10 +86,17 @@ class InpaintModelInfo(TypedDict):
family: str
name: str
description: str
backend: str # "sdxl_inpaint" | "controlnet"
backend: str # "flux_fill" | "sdxl_inpaint" | "controlnet"
INPAINT_MODELS: List[InpaintModelInfo] = [
{
"id": "flux_fill",
"family": "FLUX",
"name": "FLUX.1 Fill",
"description": "FLUX Fill 专用补全",
"backend": "flux_fill",
},
{
"id": "sdxl_inpaint",
"family": "SDXL",
@@ -100,8 +107,8 @@ INPAINT_MODELS: List[InpaintModelInfo] = [
{
"id": "controlnet",
"family": "ControlNet",
"name": "ControlNet (placeholder)",
"description": "ControlNet 补全/控制生成（当前统一封装暂未实现）。",
"name": "SD1.5 + ControlNet Inpaint",
"description": "runway inpaint + v11p inpaint ControlNet。",
"backend": "controlnet",
},
]
@@ -159,10 +166,12 @@ class DepthConfig:
@dataclass
class InpaintConfig:
# 统一补全默认后端
backend: Literal["sdxl_inpaint", "controlnet"] = "sdxl_inpaint"
# SDXL Inpaint 的基础模型（可写 HuggingFace model id 或本地目录）
sdxl_base_model: str = "stabilityai/stable-diffusion-xl-base-1.0"
# 统一补全默认后端（与 model key 对齐；实际路由以请求的 model_name 为准）
backend: Literal["flux_fill", "sdxl_inpaint", "controlnet"] = "flux_fill"
# FLUX Fill（gated：须在 Hugging Face 同意条款，并建议设置环境变量 HF_TOKEN）
flux_fill_model: str = "black-forest-labs/FLUX.1-Fill-dev"
# SDXL：须为专用 inpaint 权重（勿用 xl-base 当补全底模）
sdxl_base_model: str = "diffusers/stable-diffusion-xl-1.0-inpainting-0.1"
# ControlNet Inpaint 基础模型与 controlnet 权重
controlnet_base_model: str = "runwayml/stable-diffusion-inpainting"
controlnet_model: str = "lllyasviel/control_v11p_sd15_inpaint"
python_server/config_loader.py
+6
View File
@@ -85,6 +85,12 @@ def get_sdxl_base_model_from_app(app_cfg: AppConfig | None = None) -> str:
return app_cfg.inpaint.sdxl_base_model
def get_flux_fill_model_from_app(app_cfg: AppConfig | None = None) -> str:
if app_cfg is None:
app_cfg = load_app_config()
return app_cfg.inpaint.flux_fill_model
def get_controlnet_base_model_from_app(app_cfg: AppConfig | None = None) -> str:
if app_cfg is None:
app_cfg = load_app_config()
python_server/model/Inpaint/__init__.py
+2
View File
@@ -1 +1,3 @@
from .flux_fill.loader import make_flux_fill_predictor
__all__ = ["make_flux_fill_predictor"]
python_server/model/Inpaint/flux_fill/__init__.py
+3
View File
@@ -0,0 +1,3 @@
from .loader import make_flux_fill_predictor
__all__ = ["make_flux_fill_predictor"]
python_server/model/Inpaint/flux_fill/loader.py
+109
View File
@@ -0,0 +1,109 @@
from __future__ import annotations
import os
from typing import Any, Callable
import numpy as np
from PIL import Image
def make_flux_fill_predictor(
model_id: str,
device: str | None = None,
) -> Callable[[Image.Image, Image.Image, str, str], Image.Image]:
import torch
from diffusers import FluxFillPipeline
if device is None:
device = "cuda" if torch.cuda.is_available() else "cpu"
if device == "cuda" and torch.cuda.is_bf16_supported():
torch_dtype = torch.bfloat16
elif device == "cuda":
torch_dtype = torch.float16
else:
torch_dtype = torch.float32
hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN")
pipe = FluxFillPipeline.from_pretrained(
model_id,
torch_dtype=torch_dtype,
token=hf_token,
)
if device == "cuda":
try:
pipe.enable_model_cpu_offload()
except Exception:
try:
pipe.enable_vae_tiling()
except Exception:
pass
pipe.to("cuda")
else:
pipe.to(device)
def _predict(
image: Image.Image,
mask: Image.Image,
prompt: str,
negative_prompt: str = "",
strength: float = 1.0,
guidance_scale: float = 30.0,
num_inference_steps: int = 28,
max_side: int = 1024,
max_sequence_length: int = 512,
**_: Any,
) -> Image.Image:
_ = negative_prompt
image = image.convert("RGB")
mask = mask.convert("L")
orig_w, orig_h = image.size
run_w, run_h = orig_w, orig_h
if max(orig_w, orig_h) > max_side:
scale = max_side / float(max(orig_w, orig_h))
run_w = int(round(orig_w * scale))
run_h = int(round(orig_h * scale))
run_w = max(8, run_w - (run_w % 8))
run_h = max(8, run_h - (run_h % 8))
if (run_w, run_h) != (orig_w, orig_h):
image_run = image.resize((run_w, run_h), resample=Image.BICUBIC)
mask_run = mask.resize((run_w, run_h), resample=Image.NEAREST)
else:
image_run = image
mask_run = mask
p = (prompt or "").strip()
if not p:
p = "Chinese ink painting background, natural texture, seamless fill, no figures"
st = float(strength)
st = min(1.0, max(0.05, st))
if device == "cuda":
try:
torch.cuda.empty_cache()
except Exception:
pass
out = pipe(
prompt=p,
image=image_run,
mask_image=mask_run,
height=run_h,
width=run_w,
strength=st,
guidance_scale=float(guidance_scale),
num_inference_steps=int(num_inference_steps),
max_sequence_length=int(max_sequence_length),
).images[0]
out = out.convert("RGB")
if out.size != (orig_w, orig_h):
out = out.resize((orig_w, orig_h), resample=Image.BICUBIC)
return out
return _predict
python_server/model/Inpaint/inpaint_loader.py
+29 -5
View File
@@ -4,6 +4,7 @@ from __future__ import annotations
统一的补全（Inpaint）模型加载入口。
当前支持：
- FLUX Fill（diffusers FluxFillPipeline，如 FLUX.1-Fill-dev）
- SDXL Inpaint（diffusers AutoPipelineForInpainting）
- ControlNet（Stable Diffusion + ControlNet Inpaint）
- LaMa（torchscript big-lama，本地封装）
@@ -11,7 +12,7 @@ from __future__ import annotations
from dataclasses import dataclass
from enum import Enum
from typing import Callable
from typing import Any, Callable
import numpy as np
from PIL import Image
@@ -21,10 +22,12 @@ from config_loader import (
get_sdxl_base_model_from_app,
get_controlnet_base_model_from_app,
get_controlnet_model_from_app,
get_flux_fill_model_from_app,
)
class InpaintBackend(str, Enum):
FLUX_FILL = "flux_fill"
SDXL_INPAINT = "sdxl_inpaint"
CONTROLNET = "controlnet"
LAMA = "lama"
@@ -36,6 +39,7 @@ class UnifiedInpaintConfig:
device: str | None = None
# SDXL base model (HF id 或本地目录)，不填则用 config.py 的默认值
sdxl_base_model: str | None = None
flux_fill_model: str | None = None
@dataclass
@@ -100,7 +104,7 @@ def _make_sdxl_inpaint_predictor(
- 输出：PIL RGB 结果图
"""
import torch
from diffusers import AutoPipelineForText2Image, AutoPipelineForInpainting
from diffusers import AutoPipelineForInpainting
app_cfg = load_app_config()
base_model = cfg.sdxl_base_model or get_sdxl_base_model_from_app(app_cfg)
@@ -112,13 +116,12 @@ def _make_sdxl_inpaint_predictor(
torch_dtype = torch.float16 if device == "cuda" else torch.float32
pipe_t2i = AutoPipelineForText2Image.from_pretrained(
pipe = AutoPipelineForInpainting.from_pretrained(
base_model,
torch_dtype=torch_dtype,
variant="fp16" if device == "cuda" else None,
use_safetensors=True,
).to(device)
pipe = AutoPipelineForInpainting.from_pipe(pipe_t2i).to(device)
# 省显存设置（尽量不改变输出语义）
# 注意：CPU offload 会明显变慢，但能显著降低显存占用。
@@ -149,6 +152,7 @@ def _make_sdxl_inpaint_predictor(
guidance_scale: float = 7.5,
num_inference_steps: int = 30,
max_side: int = 1024,
**_: Any,
) -> Image.Image:
image = image.convert("RGB")
# diffusers 要求 mask 为单通道，白色区域为需要重绘
@@ -257,6 +261,7 @@ def _make_controlnet_predictor(_: UnifiedInpaintConfig):
num_inference_steps: int = 30,
controlnet_conditioning_scale: float = 1.0,
max_side: int = 768,
**_: Any,
) -> Image.Image:
import cv2
import numpy as np
@@ -316,6 +321,22 @@ def _make_controlnet_predictor(_: UnifiedInpaintConfig):
return _predict
def _make_flux_fill_predictor(cfg: UnifiedInpaintConfig):
import torch
from .flux_fill.loader import make_flux_fill_predictor as _flux_fill_factory
app_cfg = load_app_config()
model_id = cfg.flux_fill_model or get_flux_fill_model_from_app(app_cfg)
device = cfg.device
if device is None:
device = app_cfg.inpaint.device
device = "cuda" if device.startswith("cuda") and torch.cuda.is_available() else "cpu"
return _flux_fill_factory(model_id=model_id, device=device)
def _make_lama_predictor(cfg: UnifiedInpaintConfig):
import torch
@@ -340,7 +361,7 @@ def _make_lama_predictor(cfg: UnifiedInpaintConfig):
max_side: int = 1024,
**_kwargs,
) -> Image.Image:
# LaMa 不使用 prompt/negative_prompt，也不使用 strength/steps 等扩散参数
# big-lama：仅 (RGB, mask)；无文本条件。
_ = (strength, guidance_scale, num_inference_steps, max_side)
return lama(image, mask)
@@ -355,6 +376,9 @@ def build_inpaint_predictor(
"""
cfg = cfg or UnifiedInpaintConfig()
if cfg.backend == InpaintBackend.FLUX_FILL:
return _make_flux_fill_predictor(cfg), InpaintBackend.FLUX_FILL
if cfg.backend == InpaintBackend.SDXL_INPAINT:
return _make_sdxl_inpaint_predictor(cfg), InpaintBackend.SDXL_INPAINT
python_server/server.py
+4 -5
View File
@@ -1,12 +1,11 @@
"""
source .venv/bin/activate # 仓库根 hfut-bishe/.venv
cd python_server && uvicorn app.main:app --host 0.0.0.0 --port 8000
"""
from __future__ import annotations
import os
from dotenv import load_dotenv
load_dotenv()
from app.main import app
__all__ = ["app"]