FLUX Integration for Image Generation

FLUX integration for image generation

FLUX from Black Forest Labs (the team behind Stable Diffusion) is the current state-of-the-art in realistic image generation. FLUX.1 Dev and FLUX.1 Pro outperform SDXL and are comparable to Midjourney v6 in quality.

Model options

Replicate API integration

import replicate
import httpx
import asyncio

async def generate_flux(
    prompt: str,
    model: str = "flux-dev",  # flux-pro, flux-dev, flux-schnell
    aspect_ratio: str = "1:1",
    output_format: str = "webp",
    guidance: float = 3.5,
    steps: int = 28
) -> bytes:
    model_map = {
        "flux-pro": "black-forest-labs/flux-pro",
        "flux-dev": "black-forest-labs/flux-dev",
        "flux-schnell": "black-forest-labs/flux-schnell"
    }

    output = await replicate.async_run(
        model_map[model],
        input={
            "prompt": prompt,
            "aspect_ratio": aspect_ratio,
            "output_format": output_format,
            "output_quality": 90,
            "guidance": guidance,
            "num_inference_steps": steps,
        }
    )

    async with httpx.AsyncClient() as client:
        response = await client.get(str(output[0]))
        return response.content

Self-hosted с diffusers

from diffusers import FluxPipeline
import torch

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    torch_dtype=torch.bfloat16
)
pipe.enable_model_cpu_offload()  # Для GPU < 24 GB

def generate(prompt: str, width: int = 1024, height: int = 1024) -> bytes:
    import io
    image = pipe(
        prompt,
        height=height,
        width=width,
        guidance_scale=3.5,
        num_inference_steps=50,
        max_sequence_length=512,
        generator=torch.Generator("cpu").manual_seed(0)
    ).images[0]

    buf = io.BytesIO()
    image.save(buf, format="PNG")
    return buf.getvalue()

FLUX.1 Schnell for rapid prototyping

# Schnell: Apache 2.0, 4 шага вместо 50
pipe_schnell = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-schnell",
    torch_dtype=torch.bfloat16
)

image = pipe_schnell(
    prompt="professional photo of a product on white background",
    num_inference_steps=4,  # 4 шага достаточно
    guidance_scale=0.0,     # Schnell не использует CFG
).images[0]

FLUX ControlNet allows pose, depth, and edge control (like ControlNet for SD). Timeline: Replicate API integration – 1 day. Self-hosted with a queue – 1 week.