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

Overview

FLUX.1 Kontext [dev] is an open-weight, text-guided image editing model from Black Forest Labs focused on precise, context-aware edits to existing images. Unlike general text-to-image models, Kontext is built for image-to-image workflows: you provide an image plus a natural-language edit instruction and the model performs local or global edits (style transfer, object/clothing changes, text replacement, background swaps) while preserving composition and subject identity. The open-weight “dev” checkpoint is distributed under a non-commercial FLUX.1 [dev] license for research and experimentation (see the official repo and model card on GitHub / Hugging Face). (Sources: Black Forest Labs repo and Hugging Face model card.) Kontext is a 12-billion-parameter rectified-flow / flow-matching transformer designed and engineered for iterative, multi-step editing with strong character consistency and prompt-following. The project provides local inference tooling (CLI, demos, Diffusers examples), day‑one ComfyUI templates and TensorRT engine builds (BF16/FP8/FP4) for Blackwell/NVIDIA acceleration. Black Forest Labs reports that Kontext leads human-preference evaluations on their KontextBench benchmark and offers a self-serve commercial licensing path; the open weights remain available for non-commercial use and for research. (Sources: Black Forest Labs announcements, GitHub, Replicate model page, BFL licensing pages.)

Key Features

• Text-guided local and global image edits (replace objects, recolor, retouch).
• Style transfer that preserves subject identity and composition across edits.
• Text-in-image editing (signs, posters) with layout- and font-preserving results.
• Character consistency for multi-step edits preserving faces and poses without finetuning.
• Multi-image composition and in-context merging while keeping visual coherence.
• Optimized inference: TensorRT builds (BF16/FP8/FP4) and ComfyUI/Diffusers support.

Example Usage

Example (python):

## Example: run FLUX.1 Kontext [dev] using the Replicate Python client
# Requires: pip install replicate
# Set REPLICATE_API_TOKEN in your environment first

import replicate

# run a single image edit (input can be an open URL or a local file handle)
output = replicate.run(
  "black-forest-labs/flux-kontext-dev",
  input={
    "prompt": "Change the storefront sign to read 'Bistro Luna' in white serif font",
    "img_cond_path": open("./input.jpg", "rb"),
    "num_inference_steps": 30,
    "guidance": 2.5,
    "aspect_ratio": "match_input_image"
  }
)

# replicate.run returns FileOutput objects for image outputs; save the first result
with open("output_edited.jpg", "wb") as f:
    f.write(output[0].read())

# Alternative local option: use the official FLUX CLI shipped in the GitHub repo
# Example (shell): python -m flux kontext --img_cond_path ./input.jpg --prompt "..." --num_steps 30

# Sources: Replicate Python client docs and model API playground examples.

Pricing

Commercial self-hosted licensing is offered by Black Forest Labs: FLUX.1 Kontext [dev] $999/month for up to 100,000 images, then $0.01 per image beyond that. The dev weights are available under a non-commercial FLUX.1 [dev] license for research; commercial access and pricing details are published on Black Forest Labs’ licensing pages (see bfl.ai). (Sources: https://bfl.ai/licensing, https://replicate.com/black-forest-labs/flux-kontext-dev)

Benchmarks

Parameter count: 12 billion (Source: https://github.com/black-forest-labs/flux)

Replicate usage (public endpoint runs): 5.6M runs (model page total) (Source: https://replicate.com/black-forest-labs/flux-kontext-dev)

Human-preference benchmark (KontextBench): Ranked highest in reported KontextBench human-preference evaluations vs. several open and closed models (Source: https://bfl.ai/announcements/flux-1-kontext-dev)

Hardware optimizations: TensorRT engine variants available (BF16, FP8, FP4) for NVIDIA Blackwell (Source: https://bfl.ai/announcements/flux-1-kontext-dev)

Commercial license (self-serve): $999 / month for up to 100,000 images; $0.01 per image above (Source: https://bfl.ai/licensing)

Key Information

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