Flux Schnell - AI Image Models Tool

Overview

Flux Schnell (FLUX.1 [schnell]) is a fast, open-source text-to-image model from Black Forest Labs optimized for rapid iteration, local development, and personal projects. The model is a 12 billion-parameter rectified flow transformer trained with a latent adversarial diffusion distillation technique that enables high-quality image synthesis in very few sampling steps (typically 1–4). Flux Schnell is distributed under an Apache‑2.0 license and the weights are available for local use (Hugging Face), and the model is published and runnable on platforms such as Replicate and ComfyUI (for node-based local inference). (Sources: Replicate model page; Black Forest Labs model card.) Its design trades some peak fidelity for dramatic reductions in latency and compute: a specialized "go_fast" inference mode exposes an fp8-quantized, compiled variant with an optimized attention kernel for accelerated real-time-style generation. That makes Flux Schnell well suited for prototyping, interactive tools, rapid concept art, and experiments where generation speed and low resource requirements matter more than the absolute best photorealism. For developers, the project provides reference sampling code and a GitHub repository to run locally or integrate via Replicate’s API and other hosting providers. (Sources: Replicate model page; Black Forest Labs documentation.)

Key Features

• 12B-parameter rectified flow transformer for high visual capacity
• Latent adversarial diffusion distillation enabling 1–4 step sampling
• Apache-2.0 licensed weights available on Hugging Face for local use
• go_fast inference flag: fp8 quantized compiled kernel for low-latency runs
• ComfyUI node support for GUI-based local workflows and experimentation
• Reference sampling code and a GitHub repo for integration and extensions

Example Usage

Example (python):

import os
import requests

# Example: run Flux Schnell through Replicate's Predictions API.
# Replace VERSION with the model version id from the Replicate model page.
# Set REPLICATE_API_TOKEN in your environment before running.

REPLICATE_TOKEN = os.environ.get("REPLICATE_API_TOKEN")
if not REPLICATE_TOKEN:
    raise EnvironmentError("Set REPLICATE_API_TOKEN environment variable")

VERSION = "REPLACE_WITH_MODEL_VERSION_ID"  # get from https://replicate.com/black-forest-labs/flux-schnell (API tab)
payload = {
    "version": VERSION,
    "input": {
        "prompt": "A cinematic portrait of a cyborg fox, 35mm film, dramatic lighting",
        "width": 512,
        "height": 512,
        "num_outputs": 1,
        "go_fast": True  # toggle accelerated inference variant
    }
}

resp = requests.post(
    "https://api.replicate.com/v1/predictions",
    headers={
        "Authorization": f"Token {REPLICATE_TOKEN}",
        "Content-Type": "application/json",
    },
    json=payload,
)
resp.raise_for_status()
print(resp.json())

Benchmarks

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

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

Replicate runs (as listed on model page): 588.4M runs (page listing) (Source: https://replicate.com/black-forest-labs/flux-schnell)

License: Apache-2.0 (weights available for local/commercial use) (Source: https://blackforestlabs.io/flux-1-model-card/)

Accelerated inference mode: go_fast: compiled fp8 quantization + optimized attention kernel (Source: https://replicate.com/black-forest-labs/flux-schnell)

Key Information

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