FLUX.1 - AI Image Models Tool

Overview

FLUX.1 is a family of text‑to‑image generative models from Black Forest Labs that replace traditional diffusion sampling with a flow‑matching / rectified‑flow approach, producing a distinctive "flow" aesthetic while enabling much faster sampling than many diffusion models. The family includes FLUX.1 [pro] (highest quality), FLUX.1 [dev] (open-weight, guidance‑distilled), and FLUX.1 [schnell] (speed‑optimized, local‑friendly), making the suite usable for everything from research and fine‑tuning to low‑latency local inference. ([replicate.com](https://replicate.com/blog/flux-first-impressions)) Practically, the FLUX models are implemented as ~12B‑parameter rectified flow transformers and are trained with techniques such as latent adversarial diffusion distillation to achieve high image quality in very few steps. That design yields fast end‑to‑end generation (for example, FLUX.1 [schnell] can run 512×512 images in under half a second in many cases) and is released with an option for open‑source local use (Apache‑2.0 for the [schnell] variant). The suite supports fine‑tuning workflows and integrates with tooling like ComfyUI and Diffusers, enabling both rapid prototyping and production uses. ([replicate.com](https://replicate.com/black-forest-labs/flux-schnell?utm_source=openai))

Key Features

• Flow‑matching / rectified‑flow architecture for a fluid artistic aesthetic and efficient sampling.
• Three variants (pro, dev, schnell) cover quality, open‑weight research, and ultra‑fast local use.
• Speed‑optimized [schnell] runs 512×512 images in under a second at 4 steps in many setups.
• Open licensing for [schnell] (Apache‑2.0) and published weights/code for local integration.
• Trained with latent adversarial diffusion distillation to enable high quality in 1–4 steps.
• Integrations and tooling: reference code, ComfyUI support, Hugging Face / Diffusers compatibility.
• Fine‑tuning support on Replicate with fast trainer and LoRA workflows for custom concepts.

Example Usage

Example (python):

import os
import replicate

# Set REPLICATE_API_TOKEN in your environment before running
# Model slug shown here corresponds to the speed-optimized variant on Replicate.
# See model page for exact available versions and accepted input keys. ([replicate.com](https://replicate.com/black-forest-labs/flux-schnell?utm_source=openai))

os.environ.setdefault("REPLICATE_API_TOKEN", "YOUR_TOKEN_HERE")

prompt = (
    "A sunlit enchanted library with floating books and twisted root shelves, photorealistic, 35mm"
)

# Simple example using the Replicate run helper. Adjust input keys per the model's readme.
output = replicate.run(
    "black-forest-labs/flux-schnell:latest",
    input={
        "prompt": prompt,
        "width": 512,
        "height": 512,
        "num_inference_steps": 4,
        # Many FLUX endpoints support flags like `go_fast` / `guidance_scale` depending on the version
        # e.g. "go_fast": True for the accelerated fp8 compiled path. See the model page for details.
        "go_fast": True
    },
)

# `output` typically contains a URL or list of generated image URLs depending on the model/version
print(output)

Pricing

Replicate lists per‑image pricing for hosted FLUX.1 variants: FLUX.1 [pro] $0.055/image, FLUX.1 [dev] $0.030/image, FLUX.1 [schnell] $0.003/image (per Replicate blog post). For local use the [schnell] weights are open under Apache‑2.0. Please confirm current pricing on Replicate before committing — prices and billing terms can change. ([replicate.com](https://replicate.com/blog/flux-state-of-the-art-image-generation?utm_source=openai))

Benchmarks

Model size: ≈12 billion parameters (Source: https://replicate.com/black-forest-labs/flux-schnell)

512×512 latency (FLUX.1 [schnell], 4 steps): 0.29 seconds (median), P90 0.49 s (west coast US, Python client) (Source: https://replicate.com/blog/flux-is-fast-and-open-source)

1024×1024 latency (FLUX.1 [schnell], 4 steps): 0.72 seconds (median), P90 0.95 s (west coast US, Python client) (Source: https://replicate.com/blog/flux-is-fast-and-open-source)

Typical sampling steps (schnell): 1–4 inference steps (step‑distilled) (Source: https://replicate.com/black-forest-labs/flux-schnell)

Fast variant optimizations: FP8 quantized compiled variant and optimized attention kernels for accelerated inference (Source: https://replicate.com/black-forest-labs/flux-schnell)

Key Information

• Category: Image Models
• Type: AI Image Models Tool