FLUX.1 Redux [dev] - AI Image Models Tool

Overview

FLUX.1 Redux [dev] is an open-weight image-variation adapter from Black Forest Labs that produces new versions of an input image while preserving core composition, subjects, and overall style. Designed as part of the FLUX.1 Tools suite, Redux accepts a source image plus optional textual instructions to restyle or refine the image (for example, changing lighting, color grading, or subtle details) and is intended for iterative creative workflows and rapid concept exploration. (See the model card on Replicate and Black Forest Labs’ FLUX.1 Tools announcement.) Technically, Redux is an adapter that integrates with the FLUX.1-dev family (the FLUX.1 [dev] base models) and is distributed as a [dev] variant with open weights and inference code available on Hugging Face and GitHub. It is usable via the BFL API (for higher-resolution and pro variants) and via hosted endpoints such as Replicate, or locally through Diffusers/ComfyUI using the published pipelines. Licensing is split: usage of images generated through hosted endpoints (Replicate/BFL API) is allowed commercially per the hosted-service terms, while running downloaded weights locally is restricted by the FLUX [dev] non-commercial license—check the model license for details. (Sources: Replicate model page; Black Forest Labs docs; Hugging Face model card.)

Key Features

• Image-driven variation that preserves subjects, composition, and layout.
• Combine an input image with text prompts to restyle or refine visuals.
• Supports multiple aspect ratios (1:1, 16:9, 21:9, 3:2, etc.).
• Adjustable guidance scale, seed, and 1–50 inference steps for control.
• Outputs in webp, jpg, or png; batch generation up to four images.
• Open weights and inference code available under the FLUX [dev] terms.

Example Usage

Example (python):

## Minimal example using Hugging Face Diffusers-style pipelines (adapted from the HF model card)
import torch
from diffusers import FluxPriorReduxPipeline, FluxPipeline
from PIL import Image

# Load redux prior (adapter) and base FLUX pipeline
pipe_prior_redux = FluxPriorReduxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-Redux-dev",
    torch_dtype=torch.bfloat16
).to("cuda")

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    torch_dtype=torch.bfloat16
).to("cuda")

# Prepare an input image
input_image = Image.open("./source.jpg").convert("RGB")

# Run the prior/adapter to obtain conditioning from the image + optional prompt
prior_outputs = pipe_prior_redux(
    image=input_image,
    prompt="Make the scene golden-hour, warmer tones, subtle film grain",
    guidance_scale=3.0,
    num_inference_steps=28,
)

# Use the base pipeline with the prior conditioning to sample the final image(s)
results = pipe(
    prior=prior_outputs.prior,        # pipeline-specific conditioning object
    num_inference_steps=28,
    guidance_scale=3.0,
    seed=12345
)

# Save first output
results.images[0].save("./redux_variation.png")

# Notes:
# - The exact pipeline method names and return fields may vary by Diffusers release.
# - If using the hosted Replicate API, you can instead POST to the model endpoint with the 'redux_image' field.
# Sources: Hugging Face FLUX.1-Redux-dev model card and Replicate schema.

Pricing

Black Forest Labs’ public BFL API documentation and announcements list example pricing for the FLUX.1 model family (FLUX.1 [dev] commonly shown at $0.025 per image via the BFL API; FLUX.1 [pro] and FLUX1.1 [pro] have higher per-image rates). Hosted endpoints (Replicate, ComfyUI partners, BFL API) may bill separately; running downloaded FLUX.1 [dev] weights locally is governed by the FLUX [dev] non-commercial license. Users should confirm current per-run pricing and billing rules with the provider (BFL API or Replicate) before production use. Sources: Black Forest Labs BFL API announcements and related docs.

Benchmarks

Recommended denoising steps (default): 28 steps (recommended range 28–50) (Source: https://replicate.com/black-forest-labs/flux-redux-dev/versions/96b56814e57dfa601f3f524f82a2b336ef49012cda68828cb37cde66f481b7cb/api)

Default guidance scale: 3 (guidance up to 10) (Source: https://replicate.com/black-forest-labs/flux-redux-dev/versions/96b56814e57dfa601f3f524f82a2b336ef49012cda68828cb37cde66f481b7cb/api)

Supported output megapixels (BFL API pro Ultra): Up to 4 megapixels (4MP) for FLUX1.1 [pro] Ultra via BFL API (Source: https://bfl.ai/flux-1-tools/)

FLUX.1 [dev] family parameter count (base model): ~12 billion parameters (FLUX.1-dev family) (Source: https://www.aimodels.fyi/models/huggingFace/flux.1-dev-black-forest-labs)

Input/output constraints (Replicate schema): Inputs: redux_image; outputs: webp/jpg/png, aspect ratios and megapixel presets (1, 0.25) (Source: https://replicate.com/black-forest-labs/flux-redux-dev/api/schema)

Key Information

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