FLUX.1 Redux - AI Image Models Tool

Overview

FLUX.1 Redux [dev] is an adapter built to create controlled image variations from an existing image — producing subtle, high-fidelity restylings while preserving the input’s core composition. It is distributed as an adapter for the FLUX.1 family: users run a small prior pipeline (FluxPriorReduxPipeline) to extract guidance from an input image and then feed that guidance into the main FLUX pipeline to synthesize variations that stay faithful to the reference. The developer release is hosted on Hugging Face and requires accepting the FLUX.1 Dev non‑commercial license to download weights and code. ([huggingface.co](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev)) The adapter is designed to integrate into common inference toolchains: reference implementations and a diffusers-compatible pipeline are provided so you can run Redux with FluxPipeline locally or incorporate it into larger systems (for example, API-driven image+text restyling). Black Forest Labs also exposes Redux-style functionality in their BFL API (pro tier) that supports text-guided restyling and high-resolution outputs (the company highlights support for up to ~4‑megapixel outputs in their FLUX1.1 [pro] Ultra offering). The project has active community adoption (multiple Hugging Face Spaces and an official inference repository) and is intended for research and creative workflows under the FLUX.1 Dev licensing terms. ([huggingface.co](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev))

Model Statistics

• Downloads: 12,737
• Likes: 583

License: other

Model Details

What it is: FLUX.1 Redux is an adapter (not a standalone base generator) that extracts a compressed prior from an input image and supplies that prior to a FLUX.1 generator pipeline to produce image variations. The published adapter provides a FluxPriorReduxPipeline for image encoding and a standard FluxPipeline for synthesis; typical usage runs the prior pipeline on the source image, then passes the returned tensors into FluxPipeline for sampling. ([huggingface.co](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev)) Integration and runtime: Official code samples and a reference inference repo are provided (Black Forest Labs’ flux repository). The diffusers-compatible example uses bfloat16 on CUDA and demonstrates seeding, guidance_scale, and num_inference_steps parameters. Because Redux is an adapter, it doesn’t change the underlying FLUX.1 base model parameterization; its role is to provide a learned prior that improves fidelity and consistency for image-to-image variations. Model weights for the [dev] adapter are released under the FLUX.1-dev Non-Commercial License; the repo also documents commercial licensing routes via BFL’s API and licensing pages. Technical specifics about the adapter’s internal parameter count are not published in the model card (parameters: unknown). ([github.com](https://github.com/black-forest-labs/flux))

Key Features

• Adapter that extracts an image prior to produce faithful variations.
• Diffusers-compatible pipelines: FluxPriorReduxPipeline + FluxPipeline integration.
• Supports text-guided restyling when used through BFL API (image + prompt).
• Reference implementation and sampling code available in the official repo.
• Distributed under FLUX.1-dev Non‑Commercial License for research/creative use.

Example Usage

Example (python):

import torch
from diffusers import FluxPriorReduxPipeline, FluxPipeline
from diffusers.utils import load_image

# Load the prior/adapter pipeline (extracts a prior from an input image)
pipe_prior_redux = FluxPriorReduxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-Redux-dev", torch_dtype=torch.bfloat16
).to("cuda")

# Load the main FLUX.1 generator (adapter feeds into this pipeline)
pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    text_encoder=None,
    text_encoder_2=None,
    torch_dtype=torch.bfloat16,
).to("cuda")

# Prepare input image
image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/robot.png")

# Run adapter to obtain conditioning/prior
pipe_prior_output = pipe_prior_redux(image)

# Generate a variation with guidance and steps
images = pipe(
    guidance_scale=2.5,
    num_inference_steps=50,
    generator=torch.Generator("cpu").manual_seed(0),
    **pipe_prior_output,
).images

# Save first result
images[0].save("flux-dev-redux.png")

Pricing

FLUX.1 Redux [dev] is released under the FLUX.1-dev Non‑Commercial License; downloading and local use require agreeing to the model card terms on Hugging Face. Commercial/hosted access (FLUX.1 Redux [pro] behavior) is provided via Black Forest Labs’ BFL API, which Black Forest Labs lists with per-image pricing examples for their FLUX.1 family (historical example: FLUX.1 [dev] 2.5¢/img, FLUX.1 [pro] 5¢/img, FLUX1.1 [pro] 4¢/img as reported on the vendor site). For current commercial pricing and licensing terms consult BFL’s pricing/licensing pages and API docs. ([huggingface.co](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev))

Benchmarks

Hugging Face likes: 583 (Source: https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev (model card). ([huggingface.co](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev)))

Hugging Face - downloads (last month): 12,737 (Source: https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev (model card). ([huggingface.co](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev)))

Spaces / demos using adapter: 42 (Hugging Face Spaces listed) (Source: https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev (model card). ([huggingface.co](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev)))

Vendor claim — image-variation quality: Claimed state‑of‑the‑art image-variation performance (vendor benchmark) (Source: Black Forest Labs FLUX.1 Tools announcement. ([bfl.ai](https://bfl.ai/flux-1-tools/?utm_source=openai)))

Key Information

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