1
HF_TOKEN=hf_여기에_발급받은_토큰_붙여넣기

1
from fastapi import FastAPI, HTTPException, UploadFile, File
2
from fastapi.middleware.cors import CORSMiddleware
3
from pydantic import BaseModel, Field
4
import requests, os
5
from dotenv import load_dotenv

1
load_dotenv()
2

3
app = FastAPI(
4
    title="OTT 허깅페이스 무료 API 데모",
5
    version="1.0.0",
6
)
7

8
app.add_middleware(
9
    CORSMiddleware,
10
    allow_origins=["*"],
11
    allow_methods=["*"],
12
    allow_headers=["*"],
13
)
14

15
BASE = "https://router.huggingface.co/hf-inference/models"
16
VLM_URL = "https://router.huggingface.co/v1/chat/completions"
17
VLM_MODEL = "Qwen/Qwen3-VL-8B-Instruct"
18
TIMEOUT = 60

1
def get_token() -> str:
2
    t = os.getenv("HF_TOKEN")
3
    if not t:
4
        raise HTTPException(500, "HF_TOKEN 환경변수 없음")
5
    return t

1
def hf_img(model: str, data: bytes) -> dict:
2
    try:
3
        r = requests.post(
4
            f"{BASE}/{model}",
5
            headers={"Authorization": f"Bearer {get_token()}"},
6
            data=data,
7
            timeout=TIMEOUT,
8
        )
9
    except requests.exceptions.Timeout:
10
        raise HTTPException(504, "모델 응답 시간 초과")
11
    if r.status_code != 200:
12
        raise HTTPException(r.status_code, r.text)
13
    return r.json()

1
def vlm(prompt: str, img_url: str, sys_msg: str = "") -> str:
2
    msgs = []
3
    if sys_msg:
4
        msgs.append({"role": "system", "content": sys_msg})
5
    msgs.append({
6
        "role": "user",
7
        "content": [
8
            {"type": "image_url", "image_url": {"url": img_url}},
9
            {"type": "text", "text": prompt},
10
        ],
11
    })
12
    r = requests.post(
13
        VLM_URL,
14
        headers={
15
            "Authorization": f"Bearer {get_token()}",
16
            "Content-Type": "application/json",
17
        },
18
        json={"model": VLM_MODEL, "messages": msgs, "max_tokens": 400},
19
        timeout=TIMEOUT,
20
    )
21
    if r.status_code != 200:
22
        raise HTTPException(r.status_code, r.text)
23
    return r.json()["choices"][0]["message"]["content"]

1
class Text(BaseModel):
2
    text: str = Field(..., example="연출도 좋고 배우 연기가 압권이었다")
3

4
class Genre(BaseModel):
5
    text: str = Field(..., example="우주비행사들이 웜홀을 통해 새 행성을 찾아 떠난다")
6
    labels: list[str] = Field(
7
        default=["액션", "로맨스", "공포", "코미디", "SF", "드라마", "스릴러"],
8
    )
9

10
class QA(BaseModel):
11
    question: str = Field(..., example="주인공의 연인은 누구인가?")
12
    context: str = Field(..., example="Tony Stark is a genius inventor. Pepper Potts is his love interest.")
13

14
class Trans(BaseModel):
15
    text: str = Field(..., example="오늘 밤 별이 참 예쁘다")
16
    src: str = Field(default="ko", example="ko")
17
    tgt: str = Field(default="en", example="en")
18

19
class Similar(BaseModel):
20
    query: str = Field(..., example="우주를 배경으로 한 감동적인 SF 영화")
21
    candidates: list[str] = Field(..., example=[
22
        "외계인이 지구를 침공하는 액션 영화",
23
        "고등학교 첫사랑 로맨스",
24
        "화성 탐사 중 고립된 우주비행사의 생존기",
25
    ])
26

27
class ImgReq(BaseModel):
28
    image_url: str = Field(..., example="http://images.cocodataset.org/val2017/000000039769.jpg")
29
    prompt: str = Field(..., example="이 사진에 무엇이 보이는지 한국어로 설명해")

1
@app.post("/review/spoiler", tags=["리뷰"])
2
def spoiler(req: Text):
3
    r = hf("facebook/bart-large-mnli", {
4
        "inputs": req.text,
5
        "parameters": {"candidate_labels": ["spoiler", "no spoiler"]},
6
    })
7
    if isinstance(r, list) and r:
8
        top = r[0]
9
        return {"is_spoiler": top["label"] == "spoiler", "confidence": round(top["score"], 3)}
10
    if "labels" in r:
11
        return {"is_spoiler": r["labels"][0] == "spoiler", "confidence": round(r["scores"][0], 3)}
12
    return r

1
@app.post("/content/summary", tags=["콘텐츠"])
2
def summary(req: Text):
3
    r = hf("facebook/bart-large-cnn", {
4
        "inputs": req.text,
5
        "parameters": {"max_length": 80, "min_length": 20},
6
    })
7
    if isinstance(r, list) and r:
8
        return {"summary": r[0].get("summary_text", "")}
9
    return r

1
@app.post("/content/ner", tags=["콘텐츠"])
2
def ner(req: Text):
3
    r = hf("dslim/bert-base-NER", {"inputs": req.text})
4
    if isinstance(r, list):
5
        persons = list(set(
6
            x["word"] for x in r
7
            if x.get("entity_group") == "PER" and x.get("score", 0) > 0.8
8
        ))
9
        return {"persons": persons}
10
    return r

1
@app.post("/subtitle/translate", tags=["자막"])
2
def translate(req: Trans):
3
    r = hf(f"Helsinki-NLP/opus-mt-{req.src}-{req.tgt}", {"inputs": req.text})
4
    if isinstance(r, list) and r:
5
        return {"translated": r[0].get("translation_text", "")}
6
    return r

1
@app.post("/content/qa", tags=["콘텐츠"])
2
def qa(req: QA):
3
    r = hf(
4
        "google-bert/bert-large-uncased-whole-word-masking-finetuned-squad",
5
        {"inputs": {"question": req.question, "context": req.context}},
6
    )
7
    if "answer" in r:
8
        return {"answer": r["answer"], "score": round(r["score"], 3)}
9
    return r

1
@app.post("/image/detect", tags=["이미지"])
2
async def detect(file: UploadFile = File(...)):
3
    data = await file.read()
4
    r = hf_img("facebook/detr-resnet-50", data)
5
    if isinstance(r, list):
6
        return [
7
            {"label": x["label"], "score": round(x["score"], 3), "box": x["box"]}
8
            for x in r if x.get("score", 0) > 0.7
9
        ]
10
    return r

1
@app.post("/image/nsfw", tags=["이미지"])
2
async def nsfw(file: UploadFile = File(...)):
3
    data = await file.read()
4
    r = hf_img("Falconsai/nsfw_image_detection", data)
5
    if isinstance(r, list):
6
        for x in r:
7
            if x.get("label") == "nsfw":
8
                return {"is_nsfw": x["score"] > 0.5, "score": round(x["score"], 3)}
9
        return {"is_nsfw": False, "score": 0}
10
    return r

1
@app.post("/image/face", tags=["이미지"])
2
async def face(file: UploadFile = File(...)):
3
    data = await file.read()
4
    r = hf_img("jonathandinu/face-parsing", data)
5
    if isinstance(r, list):
6
        return [{"label": x["label"], "score": round(x["score"], 3)} for x in r]
7
    return r

1
@app.post("/audio/asr", tags=["오디오"])
2
async def asr(file: UploadFile = File(...)):
3
    data = await file.read()
4
    r = hf_img("openai/whisper-large-v3", data)
5
    return {"text": r.get("text", "")}

1
@app.post("/vlm/describe", tags=["멀티모달"])
2
def describe(req: ImgReq):
3
    result = vlm(req.prompt, req.image_url)
4
    return {"description": result}

1
@app.post("/vlm/ground", tags=["멀티모달"])
2
def ground(req: ImgReq):
3
    prompt = (
4
        "이 이미지에서 모든 객체를 탐지하고 바운딩 박스 좌표를 JSON 배열로 반환해. "
5
        '형식: [\{"label":"한국어이름","confidence":0.9,"bbox":[x1,y1,x2,y2]\}]'
6
    )
7
    result = vlm(prompt, req.image_url)
8
    return {"raw": result}

1
@app.post("/vlm/genre", tags=["멀티모달"])
2
def vlm_genre(req: ImgReq):
3
    prompt = (
4
        "이 이미지가 어떤 OTT 장르에 어울리는지 분류해. "
5
        "선택지: [액션, 로맨스, 힐링, 공포, SF, 자연다큐]. "
6
        'JSON으로 \{"genre":"선택","reason":"이유"\} 반환'
7
    )
8
    result = vlm(prompt, req.image_url)
9
    return {"raw": result}

1
@app.post("/vlm/cua", tags=["멀티모달"])
2
def cua(req: ImgReq):
3
    sys_msg = (
4
        "너는 모바일 앱 자동화 에이전트다. "
5
        "스크린샷을 보고 다음 액션을 JSON으로 반환한다: "
6
        '\{"action":"tap|swipe|type","target":"요소이름","coordinates":[x,y],"description":"설명"\}'
7
    )
8
    result = vlm(req.prompt, req.image_url, sys_msg=sys_msg)
9
    return {"raw": result}

1
@app.post("/vlm/ocr", tags=["멀티모달"])
2
def ocr(req: ImgReq):
3
    prompt = "이 이미지에 텍스트가 있으면 모두 읽어서 한국어로 번역해줘"
4
    result = vlm(prompt, req.image_url)
5
    return {"text": result}

1
@app.post("/vlm/ai-detect", tags=["멀티모달"])
2
def ai_detect(req: ImgReq):
3
    prompt = (
4
        "이 이미지의 모든 시각적 요소를 분석하고, "
5
        "어떤 기술로 만들어졌는지 추정해줘 "
6
        "(사진/3D렌더/일러스트/AI생성)"
7
    )
8
    result = vlm(prompt, req.image_url)
9
    return {"analysis": result}