perf(接口2/3): 混合调用零模型换出且响应<12s

- 三接口统一 ComfyUI 重绘 prompt「填充遮罩区域的头发,皮肤加一点磨皮」,
  避免 CLIP 文本条件缓存失效导致的反复重载(单卡装不下 Flux+CLIP 同驻)。
- 接口2 女重绘整条管线(swapHair+ComfyUI)送模型前限边 REDRAW_MAX_SIDE(默认1024),
  overlay 预览保持全分辨率,结果放大回原尺寸。
- 接口2 男/接口3 单段推理经 _prep_comfy_input 限边 GROW_B_MAX_SIDE(默认1024)。
- 修复真实大图(1257x1495)全分辨率送模型导致 13~21s 且把 Flux 挤出显存的问题。
- comfyui.py 增加输入尺寸日志;service.py 增加女重绘分段计时(swap/blend/ComfyUI)。

实测真实图三接口任意交替: 女9~12s / 男7.8s / 接口3 6.6s,CLIP/Flux 重载 0 次。

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
xsl
2026-07-19 20:08:21 +08:00
co-authored by Cursor
parent 08b31a3baa
commit 2993a0b948
6 changed files with 219 additions and 11 deletions
+127 -5
View File
@@ -40,6 +40,18 @@ _REPO = os.path.dirname(os.path.dirname(__file__))
_TEXTURE_DIR = os.path.join(_REPO, "hairline_texture")
_BLACK_TEXTURE_DIR = os.path.join(_REPO, "hairline_texture_black")
# 三接口(接口2女重绘 / 接口2男 / 接口3)统一的 ComfyUI 重绘 prompt。
# 关键:ComfyUI 单卡显存装不下 Flux(7.7G)+qwen CLIP(3.9G) 同驻,靠缓存 CLIP 文本条件避免重载。
# prompt 不同会使缓存失效 → 重载 CLIP 并挤出 Flux(每次 +4s)。三接口用同一字符串即可全程命中。
# 与 app.py 接口2/接口3 的默认 prompt 保持一致;可用 REDRAW_PROMPT 覆盖。
_REDRAW_PROMPT = os.getenv("REDRAW_PROMPT", "填充遮罩区域的头发,皮肤加一点磨皮")
# 接口2 女重绘整条管线(swapHair + ComfyUI)送模型前限边。真实照片常达 1257x1495:
# 全分辨率 ComfyUI 重绘要 13~21s 且激活显存把模型挤出。女性路径含 swapHair(SD WebUI ~5.3s
# 固定地板) + ComfyUI 两段串行。用 1024 优先画质(女性总耗时 ~11~12s,偶尔踩 12s);
# 若要更稳的 <12s 余量可设 REDRAW_MAX_SIDE=896ComfyUI ~4s,女性 9~11s)。
_REDRAW_MAX_SIDE = int(os.getenv("REDRAW_MAX_SIDE", "1024"))
def _call_local_redraw(image_png_bytes, mask_png_bytes, timeout=300.0):
"""直接调 ComfyUI 重绘(替代原 local_test HTTP 服务)。
@@ -47,7 +59,27 @@ def _call_local_redraw(image_png_bytes, mask_png_bytes, timeout=300.0):
失败抛异常(调用方负责 try/except 跳过)。
"""
from .redraw import run_redraw
return run_redraw(image_png_bytes, mask_png_bytes, timeout=timeout)
img = cv2.imdecode(np.frombuffer(image_png_bytes, np.uint8), cv2.IMREAD_UNCHANGED)
scale = 1.0
orig_w = orig_h = 0
if img is not None:
orig_h, orig_w = img.shape[:2]
m = max(orig_h, orig_w)
if _REDRAW_MAX_SIDE > 0 and m > _REDRAW_MAX_SIDE:
scale = _REDRAW_MAX_SIDE / float(m)
nw, nh = max(1, round(orig_w * scale)), max(1, round(orig_h * scale))
msk = cv2.imdecode(np.frombuffer(mask_png_bytes, np.uint8), cv2.IMREAD_UNCHANGED)
img_s = cv2.resize(img, (nw, nh), interpolation=cv2.INTER_AREA)
msk_s = cv2.resize(msk, (nw, nh), interpolation=cv2.INTER_NEAREST)
image_png_bytes = cv2.imencode(".png", img_s)[1].tobytes()
mask_png_bytes = cv2.imencode(".png", msk_s)[1].tobytes()
logger.info("接口2女 缩图送 Comfy: %dx%d%dx%d (max_side=%d)",
orig_w, orig_h, nw, nh, _REDRAW_MAX_SIDE)
out = run_redraw(image_png_bytes, mask_png_bytes, timeout=timeout,
prompt=_REDRAW_PROMPT)
if scale < 1.0 and out:
out = _upscale_png_to(out, orig_w, orig_h)
return out
# 发际线贴图档位:middle=默认(hairline_texture/)high/low 各自独立文件夹。
_TEXTURE_DIRS = {
@@ -190,9 +222,12 @@ def generate_grow_results(image_bgr: np.ndarray, gender: str, use_mask: bool = T
if not use_mask:
try:
h, w = image_bgr.shape[:2]
img_s, msk_s, gsc = _prep_comfy_input(image_bgr, np.zeros((h, w), np.uint8))
buf = io.BytesIO()
compose_comfy_rgba(image_bgr, np.zeros((h, w), np.uint8)).save(buf, format="PNG", compress_level=1)
compose_comfy_rgba(img_s, msk_s).save(buf, format="PNG", compress_level=1)
shared_grown = comfyui.run(buf.getvalue(), prompt=prompt, workflow_path=workflow_path)
if gsc < 1.0 and shared_grown:
shared_grown = _upscale_png_to(shared_grown, w, h)
except Exception as e: # noqa: BLE001
logger.warning("接口2 生发图失败(无遮罩)%s", e)
@@ -210,9 +245,12 @@ def generate_grow_results(image_bgr: np.ndarray, gender: str, use_mask: bool = T
black = load_texture_rgba(_black_texture_path(white_path))
marked, mask = build_inpaint_mask(
image_bgr, ctx["landmarks"], ctx["parse_map"], ctx["points"], black)
m_s, msk_s, gsc = _prep_comfy_input(marked, mask)
buf = io.BytesIO()
compose_comfy_rgba(marked, mask).save(buf, format="PNG", compress_level=1)
compose_comfy_rgba(m_s, msk_s).save(buf, format="PNG", compress_level=1)
grown_png = comfyui.run(buf.getvalue(), prompt=prompt, workflow_path=workflow_path)
if gsc < 1.0 and grown_png:
grown_png = _upscale_png_to(grown_png, w, h)
except Exception as e: # noqa: BLE001 单张失败不拖垮整请求
logger.warning("接口2 生发图失败 type=%s%s", key, e)
@@ -253,6 +291,20 @@ def generate_grow_results_swap(image_bgr: np.ndarray, hair_styles: list[int] | N
results = []
h, w = image_bgr.shape[:2]
# 重绘管线(swapHair + ComfyUI)统一降分辨率:真实照片 swap(SD WebUI)~5s、blend、ComfyUI
# 均随分辨率线性下降。overlay 预览仍用全分辨率;grown_png 最后放大回原尺寸。
redraw_img = image_bgr
hair_mask_redraw = hair_mask_reuse
if _REDRAW_MAX_SIDE > 0 and max(h, w) > _REDRAW_MAX_SIDE:
redraw_img, _rs = _downscale_max_side(image_bgr, _REDRAW_MAX_SIDE)
_nh, _nw = redraw_img.shape[:2]
if hair_mask_redraw is not None:
hair_mask_redraw = cv2.resize(hair_mask_reuse.astype(np.uint8), (_nw, _nh),
interpolation=cv2.INTER_NEAREST).astype(bool)
logger.info("接口2女 管线降分辨率: %dx%d%dx%d (max_side=%d)",
w, h, _nw, _nh, _REDRAW_MAX_SIDE)
for order, (key, white_path) in items:
white = load_texture_rgba(white_path)
overlay = build_overlay_layer(h, w, ctx["points"], ext_faces, uv, white)
@@ -263,7 +315,10 @@ def generate_grow_results_swap(image_bgr: np.ndarray, hair_styles: list[int] | N
logger.warning("接口2 换发型:female key=%s 无对应 chang_id,跳过生发图", key)
else:
try:
data = generate_hairline_redraw(image_bgr, chang_id, hair_mask=hair_mask_reuse, **redraw_defaults)
import time as _t
_ts0 = _t.perf_counter()
data = generate_hairline_redraw(redraw_img, chang_id, hair_mask=hair_mask_redraw, **redraw_defaults)
_ts1 = _t.perf_counter()
steps = data.get("steps") or {}
# ④ final(接缝融合基底)+ ⑤-② 纯红遮罩 PNG
final_b64 = steps.get("final_base64") or ""
@@ -280,9 +335,19 @@ def generate_grow_results_swap(image_bgr: np.ndarray, hair_styles: list[int] | N
final_bytes = base64.b64decode(final_b64)
mask_bytes = base64.b64decode(mask_b64)
# 后端直接调 ComfyUI 重绘,返回重绘后的 PNG
_tr0 = _t.perf_counter()
grown_png = _call_local_redraw(final_bytes, mask_bytes)
_tr1 = _t.perf_counter()
_tm = data.get("timings_ms") or {}
logger.info("接口2女 分段计时 type=%s: swapHair管线=%.2fs (mask=%dms swap=%dms blend=%dms), ComfyUI重绘=%.2fs",
key, _ts1 - _ts0,
_tm.get("mask", 0), _tm.get("swap", 0), _tm.get("blend", 0),
_tr1 - _tr0)
if grown_png is None:
logger.warning("接口2 换发型:type=%s 重绘结果为空", key)
elif redraw_img is not image_bgr:
# 管线在降分辨率图上跑,结果放大回原尺寸
grown_png = _upscale_png_to(grown_png, w, h)
except NoFaceError:
logger.warning("接口2 换发型:type=%s 未检出人脸", key)
except Exception as e: # noqa: BLE001 单张失败不拖垮整请求
@@ -377,18 +442,73 @@ def generate_hairline_pngs(image_bgr: np.ndarray, gender: str,
return {"images": images, "best_centers": best_centers}
# 接口3 送 ComfyUI 前限边,降低峰值显存,避免与接口2 切换时把 Flux 挤出。
# 统一 prompt 后 Flux 不再被 CLIP 挤出,接口3 可用较高分辨率。可用 GROW_B_MAX_SIDE 覆盖。
_GROW_B_MAX_SIDE = int(os.getenv("GROW_B_MAX_SIDE", "1024"))
def _downscale_max_side(img_bgr: np.ndarray, max_side: int) -> tuple[np.ndarray, float]:
"""长边超过 max_side 时等比例缩小;返回 (图, scale)scale=新/旧。"""
h, w = img_bgr.shape[:2]
m = max(h, w)
if max_side <= 0 or m <= max_side:
return img_bgr, 1.0
scale = max_side / float(m)
nw = max(1, int(round(w * scale)))
nh = max(1, int(round(h * scale)))
out = cv2.resize(img_bgr, (nw, nh), interpolation=cv2.INTER_AREA)
return out, scale
def _upscale_png_to(png_bytes: bytes, out_w: int, out_h: int) -> bytes:
"""把 Comfy 输出 PNG 双线性拉回原图尺寸(仅展示对齐,不增加推理细节)。"""
arr = np.frombuffer(png_bytes, np.uint8)
img = cv2.imdecode(arr, cv2.IMREAD_UNCHANGED)
if img is None:
return png_bytes
if img.shape[1] == out_w and img.shape[0] == out_h:
return png_bytes
resized = cv2.resize(img, (out_w, out_h), interpolation=cv2.INTER_LINEAR)
ok, buf = cv2.imencode(".png", resized)
return buf.tobytes() if ok else png_bytes
def _prep_comfy_input(img_bgr: np.ndarray, mask: np.ndarray) -> tuple[np.ndarray, np.ndarray, float]:
"""单段 ComfyUI 生发(接口2男 / 接口3)送图前限边到 GROW_B_MAX_SIDE。
返回 (缩后图, 缩后遮罩, scale)scale<1 时调用方需把结果放大回原尺寸。"""
h, w = img_bgr.shape[:2]
if _GROW_B_MAX_SIDE <= 0 or max(h, w) <= _GROW_B_MAX_SIDE:
return img_bgr, mask, 1.0
out, scale = _downscale_max_side(img_bgr, _GROW_B_MAX_SIDE)
nh, nw = out.shape[:2]
msk = cv2.resize(mask, (nw, nh), interpolation=cv2.INTER_NEAREST)
logger.info("接口2男/接口3 缩图送 Comfy: %dx%d%dx%d (max_side=%d)",
w, h, nw, nh, _GROW_B_MAX_SIDE)
return out, msk, scale
def generate_grow_b(marked_bgr: np.ndarray, use_mask: bool = True, prompt: str = None):
"""接口3:检测医生手绘发际线 → 遮罩 → 送 ComfyUI 生发(仅需划线图一张)。
检测路径只用来**建遮罩**ComfyUI 输入图用 **marked 原图**(含医生手绘线,
工作流提示词会清除黑线再生发)。
进 Comfy 前若长边 > GROW_B_MAX_SIDE(默认 896)会先等比例缩小,降低峰值显存;
输出再拉回原图尺寸。
use_mask(默认 True):是否启用自动检测的遮罩,用于测试对比。
- True:检测手绘线 → 建遮罩 → alpha=255−mask(透明区=重绘区,节点44 画黄色参考区)。
- False:跳过检测,直接送划线图,alpha 全 255(空遮罩,节点26 mask 为空),
模型仅凭医生黑线参考生发。无需改工作流,唯一变量是遮罩。
Returns: {"grown_png": bytes 或 None, "status": "ok"|"no_face"|"no_line"}。
"""
orig_h, orig_w = marked_bgr.shape[:2]
marked_bgr, _scale = _downscale_max_side(marked_bgr, _GROW_B_MAX_SIDE)
if _scale < 1.0:
logger.info(
"接口3 缩图送 Comfy: %dx%d%dx%d (max_side=%d)",
orig_w, orig_h, marked_bgr.shape[1], marked_bgr.shape[0], _GROW_B_MAX_SIDE,
)
h, w = marked_bgr.shape[:2]
if use_mask:
rgb = cv2.cvtColor(marked_bgr, cv2.COLOR_BGR2RGB)
@@ -406,8 +526,10 @@ def generate_grow_b(marked_bgr: np.ndarray, use_mask: bool = True, prompt: str =
mask = np.zeros((h, w), np.uint8) # 空遮罩:alpha 全 255,跳过检测
buf = io.BytesIO()
compose_comfy_rgba(marked_bgr, mask).save(buf, format="PNG", compress_level=1) # marked 原图 + 遮罩
compose_comfy_rgba(marked_bgr, mask).save(buf, format="PNG", compress_level=1) # marked + 遮罩
grown_png = comfyui.run(buf.getvalue(), prompt=prompt)
if _scale < 1.0 and grown_png:
grown_png = _upscale_png_to(grown_png, orig_w, orig_h)
return {"grown_png": grown_png, "status": "ok"}