replicate-integration（复制整合）

Replicate API Integration

Purpose

Deploy and run AI models in production using Replicate's cloud platform. Specialized for image generation with Flux Dev, SDXL, custom LoRA fine-tuning, and prediction polling patterns.

When to Use

• Generating images with Flux Dev, SDXL, or custom models
• Fine-tuning models with custom LoRA weights
• Running predictions with polling/webhook patterns
• Deploying custom models to production
• Managing long-running AI workloads

Architecture Pattern

Project Structure

backend/
├── services/
│   ├── replicate_service.py    # Main Replicate client
│   └── model_service.py        # Model-specific logic
├── models/
│   └── replicate_models.py     # Pydantic models
├── config/
│   └── replicate_config.py     # Configuration
└── utils/
    └── polling.py              # Polling utilities

Installation

pip install replicate httpx python-dotenv pydantic

Environment Setup

# .env
REPLICATE_API_TOKEN=r8_...
FRONTEND_URL=http://localhost:3000
DEFAULT_MODEL=black-forest-labs/flux-dev
LORA_MODEL_ID=your-username/your-model-id

Quick Start

Basic Image Generation

import replicate
import os

client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))

# Simple prediction
output = client.run(
    "black-forest-labs/flux-dev",
    input={
        "prompt": "A serene mountain landscape at sunset",
        "num_outputs": 1,
        "aspect_ratio": "16:9"
    }
)

print(f"Generated image: {output[0]}")

Async Pattern with Polling

import replicate
import asyncio
from typing import List

async def generate_images_async(
    prompt: str,
    num_images: int = 4
) -> List[str]:
    client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))

    # Start prediction
    prediction = client.predictions.create(
        version="black-forest-labs/flux-dev",
        input={
            "prompt": prompt,
            "num_outputs": num_images,
            "guidance_scale": 3.5,
            "num_inference_steps": 28
        }
    )

    # Poll for completion
    while prediction.status not in ["succeeded", "failed", "canceled"]:
        await asyncio.sleep(1)
        prediction = client.predictions.get(prediction.id)

    if prediction.status == "succeeded":
        return prediction.output
    else:
        raise Exception(f"Prediction failed: {prediction.error}")

Flux Dev Integration

Complete Flux Dev Configuration

from pydantic import BaseModel, Field
from typing import Literal

class FluxDevInput(BaseModel):
    prompt: str = Field(description="Text description of image to generate")
    aspect_ratio: Literal["1:1", "16:9", "21:9", "3:2", "2:3", "4:5", "5:4", "9:16", "9:21"] = "1:1"
    num_outputs: int = Field(ge=1, le=4, default=1)
    num_inference_steps: int = Field(ge=1, le=50, default=28)
    guidance_scale: float = Field(ge=0, le=10, default=3.5)
    output_format: Literal["webp", "jpg", "png"] = "webp"
    output_quality: int = Field(ge=0, le=100, default=80)
    seed: int | None = None
    disable_safety_checker: bool = False

async def generate_flux_images(input: FluxDevInput) -> List[str]:
    client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))

    output = await client.async_run(
        "black-forest-labs/flux-dev",
        input=input.model_dump()
    )

    return output

Flux Dev Best Practices

# Optimal settings for portraits
PORTRAIT_CONFIG = {
    "aspect_ratio": "2:3",
    "num_inference_steps": 30,
    "guidance_scale": 4.0,
    "output_format": "webp",
    "output_quality": 90
}

# Optimal settings for landscapes
LANDSCAPE_CONFIG = {
    "aspect_ratio": "16:9",
    "num_inference_steps": 28,
    "guidance_scale": 3.5,
    "output_format": "webp",
    "output_quality": 85
}

# Batch generation pattern
async def batch_generate(prompts: List[str]) -> List[List[str]]:
    tasks = [generate_flux_images(FluxDevInput(prompt=p)) for p in prompts]
    return await asyncio.gather(*tasks)

LoRA Fine-Tuning Integration

Using Custom LoRA Models

class LoRAInput(BaseModel):
    prompt: str
    lora_scale: float = Field(ge=0, le=1, default=1.0, description="LoRA influence strength")
    trigger_word: str | None = Field(default=None, description="Special token for identity")
    num_outputs: int = Field(ge=1, le=10, default=1)

async def generate_with_lora(
    model_id: str,  # e.g., "daniel-carreon/danielcarrong"
    input: LoRAInput
) -> List[str]:
    client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))

    # Ensure trigger word is in prompt
    prompt = input.prompt
    if input.trigger_word and input.trigger_word not in prompt:
        prompt = f"{input.trigger_word} {prompt}"

    output = await client.async_run(
        model_id,
        input={
            "prompt": prompt,
            "lora_scale": input.lora_scale,
            "num_outputs": input.num_outputs,
            "num_inference_steps": 28,
            "guidance_scale": 3.5
        }
    )

    return output

Training Custom LoRA

async def train_lora(
    images_zip_url: str,
    trigger_word: str,
    steps: int = 1000
) -> str:
    """Train custom LoRA model on Replicate"""
    client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))

    training = client.trainings.create(
        version="ostris/flux-dev-lora-trainer",
        input={
            "input_images": images_zip_url,
            "trigger_word": trigger_word,
            "steps": steps,
            "lora_rank": 16,
            "optimizer": "adamw8bit",
            "batch_size": 1,
            "learning_rate": 4e-4,
            "caption_prefix": f"a photo of {trigger_word}"
        },
        destination=f"{os.getenv('REPLICATE_USERNAME')}/my-lora-model"
    )

    # Wait for training completion
    while training.status not in ["succeeded", "failed", "canceled"]:
        await asyncio.sleep(30)
        training = client.trainings.get(training.id)

    if training.status == "succeeded":
        return training.output  # Model URL
    else:
        raise Exception(f"Training failed: {training.error}")

Polling Patterns

Robust Polling with Retry

import asyncio
from typing import Callable, Any

class PollConfig(BaseModel):
    max_wait: int = 300  # 5 minutes
    poll_interval: float = 1.0  # 1 second
    timeout_multiplier: float = 1.5  # Backoff factor

async def poll_prediction(
    prediction_id: str,
    on_progress: Callable[[float], None] | None = None
) -> Any:
    """Poll Replicate prediction with exponential backoff"""
    client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))

    start_time = asyncio.get_event_loop().time()
    poll_interval = 1.0

    while True:
        prediction = client.predictions.get(prediction_id)

        # Report progress
        if on_progress and hasattr(prediction, 'logs'):
            progress = extract_progress(prediction.logs)
            on_progress(progress)

        # Check status
        if prediction.status == "succeeded":
            return prediction.output
        elif prediction.status in ["failed", "canceled"]:
            raise Exception(f"Prediction {prediction.status}: {prediction.error}")

        # Timeout check
        elapsed = asyncio.get_event_loop().time() - start_time
        if elapsed > 300:  # 5 minutes
            raise TimeoutError(f"Prediction timeout after {elapsed}s")

        # Exponential backoff
        await asyncio.sleep(poll_interval)
        poll_interval = min(poll_interval * 1.5, 5.0)

def extract_progress(logs: str) -> float:
    """Extract progress from logs (0.0 to 1.0)"""
    # Example: "Progress: 50%"
    import re
    match = re.search(r"Progress: (\d+)%", logs or "")
    return float(match.group(1)) / 100 if match else 0.0

Webhook Pattern (Production)

from fastapi import FastAPI, Request

app = FastAPI()

@app.post("/webhooks/replicate")
async def handle_webhook(request: Request):
    """Handle Replicate webhook callback"""
    payload = await request.json()

    prediction_id = payload["id"]
    status = payload["status"]

    if status == "succeeded":
        output = payload["output"]
        # Process completed prediction
        await save_results(prediction_id, output)
    elif status == "failed":
        error = payload["error"]
        # Handle error
        await log_error(prediction_id, error)

    return {"status": "received"}

# Start prediction with webhook
def create_prediction_with_webhook(prompt: str) -> str:
    client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))

    prediction = client.predictions.create(
        version="black-forest-labs/flux-dev",
        input={"prompt": prompt},
        webhook=f"{os.getenv('BACKEND_URL')}/webhooks/replicate",
        webhook_events_filter=["completed"]
    )

    return prediction.id

Error Handling

Comprehensive Error Handling

from enum import Enum

class ReplicateError(Exception):
    """Base Replicate error"""
    pass

class RateLimitError(ReplicateError):
    """Rate limit exceeded"""
    pass

class ModelNotFoundError(ReplicateError):
    """Model not found"""
    pass

async def safe_replicate_call(
    model: str,
    input: dict,
    max_retries: int = 3
) -> Any:
    """Call Replicate with retry logic"""
    client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))

    for attempt in range(max_retries):
        try:
            output = await client.async_run(model, input=input)
            return output
        except replicate.exceptions.ModelError as e:
            if "not found" in str(e).lower():
                raise ModelNotFoundError(f"Model {model} not found")
            raise
        except replicate.exceptions.ReplicateError as e:
            if "rate limit" in str(e).lower():
                if attempt < max_retries - 1:
                    await asyncio.sleep(2 ** attempt)  # Exponential backoff
                    continue
                raise RateLimitError("Rate limit exceeded")
            raise ReplicateError(f"Replicate error: {e}")
        except Exception as e:
            if attempt < max_retries - 1:
                await asyncio.sleep(1)
                continue
            raise

FastAPI Integration

Complete Replicate Endpoint

from fastapi import FastAPI, HTTPException, BackgroundTasks
from pydantic import BaseModel

app = FastAPI()

class GenerateRequest(BaseModel):
    prompt: str
    num_images: int = 4
    use_lora: bool = False
    lora_model_id: str | None = None

class GenerateResponse(BaseModel):
    prediction_id: str
    status: str
    images: List[str] | None = None

@app.post("/generate", response_model=GenerateResponse)
async def generate_images(request: GenerateRequest):
    """Generate images using Replicate"""
    try:
        client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))

        # Select model
        model = request.lora_model_id if request.use_lora else "black-forest-labs/flux-dev"

        # Create prediction
        prediction = client.predictions.create(
            version=model,
            input={
                "prompt": request.prompt,
                "num_outputs": request.num_images
            }
        )

        return GenerateResponse(
            prediction_id=prediction.id,
            status=prediction.status
        )
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

@app.get("/predictions/{prediction_id}")
async def get_prediction(prediction_id: str):
    """Check prediction status"""
    try:
        client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))
        prediction = client.predictions.get(prediction_id)

        return GenerateResponse(
            prediction_id=prediction.id,
            status=prediction.status,
            images=prediction.output if prediction.status == "succeeded" else None
        )
    except Exception as e:
        raise HTTPException(status_code=404, detail="Prediction not found")

Rate Limiting

Rate Limit Management

from collections import deque
from datetime import datetime, timedelta

class RateLimiter:
    def __init__(self, max_requests: int = 50, window: int = 60):
        self.max_requests = max_requests
        self.window = timedelta(seconds=window)
        self.requests: deque[datetime] = deque()

    async def acquire(self):
        """Wait if rate limit reached"""
        now = datetime.now()

        # Remove old requests
        while self.requests and self.requests[0] < now - self.window:
            self.requests.popleft()

        # Check limit
        if len(self.requests) >= self.max_requests:
            wait_time = (self.requests[0] + self.window - now).total_seconds()
            if wait_time > 0:
                await asyncio.sleep(wait_time)

        self.requests.append(now)

# Usage
limiter = RateLimiter(max_requests=50, window=60)

async def generate_with_limit(prompt: str):
    await limiter.acquire()
    return await generate_flux_images(FluxDevInput(prompt=prompt))

Best Practices

1. Always use async/await for non-blocking I/O
2. Implement polling with backoff to avoid rate limits
3. Use webhooks in production for long-running tasks
4. Cache prediction results to avoid redundant API calls
5. Monitor costs - log prediction IDs and metrics
6. Handle errors gracefully with retry logic
7. Use typed inputs with Pydantic models
8. Set timeouts for all predictions
9. Validate outputs before returning to users
10. Store metadata for debugging and analytics

Common Pitfalls

❌ Don't: Poll too frequently (wastes API calls) ✅ Do: Use exponential backoff (1s → 1.5s → 2.25s →...)

❌ Don't: Forget to handle prediction failures ✅ Do: Check status and handle errors

❌ Don't: Hardcode model versions ✅ Do: Use environment variables for flexibility

❌ Don't: Block on predictions in API endpoints ✅ Do: Return prediction ID immediately, poll separately

Complete Example: Production Service

from fastapi import FastAPI, BackgroundTasks
import replicate
from pydantic import BaseModel
from typing import List
import asyncio

app = FastAPI()

class ImageGenerationService:
    def __init__(self):
        self.client = replicate.Client(api_token=os.getenv("REPLICATE_API_TOKEN"))
        self.rate_limiter = RateLimiter(max_requests=50, window=60)

    async def generate(
        self,
        prompt: str,
        num_images: int = 4,
        model_id: str = "black-forest-labs/flux-dev"
    ) -> List[str]:
        """Generate images with rate limiting and error handling"""
        await self.rate_limiter.acquire()

        try:
            prediction = self.client.predictions.create(
                version=model_id,
                input={
                    "prompt": prompt,
                    "num_outputs": num_images,
                    "num_inference_steps": 28,
                    "guidance_scale": 3.5
                }
            )

            # Poll for completion
            output = await poll_prediction(
                prediction.id,
                on_progress=lambda p: print(f"Progress: {p*100:.0f}%")
            )

            return output
        except Exception as e:
            print(f"Generation failed: {e}")
            raise

# Global service instance
service = ImageGenerationService()

@app.post("/api/generate")
async def generate_endpoint(request: GenerateRequest):
    try:
        images = await service.generate(
            prompt=request.prompt,
            num_images=request.num_images
        )
        return {"status": "success", "images": images}
    except Exception as e:
        return {"status": "error", "message": str(e)}

Resources

• Replicate Docs
• Flux Dev Model
• LoRA Training Guide
• Python Client