# FLUX.1 Kontext Dev: Open Weights and Local Image Editing Guide

> Published 2026-04-04, updated 2026-09-05 · https://www.promptzone.com/arne_suzuki/flux-kontext-open-weight-ai-model-released-1hib

---
title: "FLUX.1 Kontext Dev: Open Weights and Local Image Editing Guide"
published: true
description: "Run FLUX.1 Kontext dev for local image editing, understand its weight license, and evaluate changes against your source image."
tags: ai,imagegeneration,comfyui,flux
---
FLUX.1 Kontext dev is Black Forest Labs’ downloadable 12B image transformer for editing images with text instructions. To run it locally, obtain access to its Hugging Face weights and load them through a supported implementation such as Diffusers or ComfyUI. [Model card][card], [Hugging Face access][hf]

The dev release provides weights for local workflows; Kontext pro and max have separate hosted access. [Release][launch], [Official repository][repo]

## What are the key facts about FLUX.1 Kontext dev?

| Field | Verified detail |
| --- | --- |
| Developer | Black Forest Labs. [Release][launch] |
| Released | Public dev weights: June 26, 2025. [Release][launch] |
| Type | Rectified-flow transformer for instruction-based image editing. [Model card][card] |
| Size or parameters | 12 billion for the dev image transformer. [Model card][card] |
| License and access | Downloadable dev weights under the FLUX.1 dev non-commercial license; separate commercial licensing is available. [Release][launch], [License][license] |
| Where it runs | Local reference code, Diffusers, or ComfyUI; hosted model access is available separately. [Model card][card] |

## What can you edit with FLUX.1 Kontext dev?

Kontext dev takes an existing image and an instruction describing the desired change. Its model card identifies character, object, and style references as supported uses without extra fine-tuning. [Model card][card]

This fits a workflow in which you already have the composition but need to alter an element. Start with a small change so that you can evaluate both the alteration and the content that should remain.

For example, use a photograph of a desk and ask to recolor its chair. Write down the chair frame, desk placement, lighting, and background as preserved features before reviewing the result.

BFL also describes iterative editing with limited visual drift as a model capability. Treat that as a reason to evaluate an edit sequence, while checking each accepted image against the starting point. [Model card][card]

A useful test might progress from upholstery color to wall color, then a different lighting condition. Keep intermediate outputs so you can identify which instruction introduced an unwanted change.

For the family’s broader access options, read the sibling [Kontext overview](/riya_ahmadi/flux-kontext-ai-model-debuts-3988). This page focuses on the downloadable dev checkpoint.

## What are the limits of local Kontext dev editing?

Open weights and unrestricted use are different questions. The dev license governs the model and its derivatives, while BFL separately describes permitted uses of generated outputs. [License][license]

The model card states that outputs can be used for commercial purposes as described in that license. Evaluate the actual deployment separately from deciding how to use a finished image. [Model card][card]

BFL’s launch post provides a separate commercial licensing route. Use that route when the intended model use exceeds the terms of the downloadable release. [Release][launch]

Memory requirements depend on the complete pipeline. Diffusers documents quantization and offloading techniques for FLUX, rather than a universal hardware guarantee for all settings and components. [Diffusers documentation][diffusers]

Image preservation also needs review. The authors’ technical work studies consistency during iterative editing; a model’s ability to make successive edits does not establish pixel-for-pixel preservation. [Technical report][report]

Inspect the original and edited image side by side. Pay attention to faces, text, object geometry, and background details that were not part of the requested change.

BFL now recommends FLUX.2 for new editing projects in its hosted Kontext documentation. That recommendation is useful when choosing a new pipeline, while Kontext dev remains a separately documented local model. [Editing documentation][editing]

## How do you run FLUX.1 Kontext dev locally?

Open the official Kontext dev repository on Hugging Face and complete its access requirements. Install compatible PyTorch and Diffusers packages using the library’s FLUX documentation. [Hugging Face card][hf], [Diffusers documentation][diffusers]

The example below uses the documented `FluxKontextPipeline`. It assumes a CUDA-capable environment with enough memory for the selected configuration and an existing local image named `chair.png`. [Diffusers documentation][diffusers]

```python
import torch
from diffusers import FluxKontextPipeline
from diffusers.utils import load_image

pipe = FluxKontextPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-Kontext-dev",
    torch_dtype=torch.bfloat16,
).to("cuda")
image = pipe(
    image=load_image("chair.png"),
    prompt="Change the chair upholstery to green; keep the room unchanged",
    guidance_scale=2.5,
).images[0]
image.save("chair-green.png")
```

The input filename and edit are illustrative. Use a picture you can inspect clearly, and keep the original file separate from the generated result.

For a first test, choose a single visible alteration. Record the requested edit, source image, model identifier, and settings; approve the output only after comparing the whole image to the source.

If you prefer a graph interface, ComfyUI publishes a native Kontext dev workflow. Follow that guide’s model links and component instructions rather than copying the setup for a different FLUX generation. [ComfyUI guide][comfy]

The [ComfyUI pillar](/tomas_novak/comfyui-2026-the-complete-guide-to-power-user-ai-image-generation-1g17) provides background on loading and organizing workflows. Save the original template before changing its structure.

For successive edits, create a record with the input image, instruction, output, and review decision at each stage. Keep a route back to the original when an intermediate result has already changed too much.

When an edit fails, separate three questions: did the pipeline execute, did the requested change happen, and did the rest of the image remain acceptable? They call for different corrective actions.

A runtime error requires checking the environment and loaded components. An incorrect recoloring may call for a clearer instruction; an acceptable recoloring with damaged lettering may call for a different editing method.

Use a small evaluation set that reflects your intended work. Include a clean object, a cluttered scene, and a picture where identity or lettering matters, then note which failure types recur.

## How does Kontext dev compare with hosted Kontext and FLUX.2?

| Option | Main distinction |
| --- | --- |
| Kontext dev | Downloadable image-editing checkpoint with a non-commercial weight license. [Model card][card] |
| Kontext pro or max | Separate hosted offerings in BFL’s model lineup, without released local weights. [Official repository][repo] |
| FLUX.2 dev | Downloadable 32B image transformer with generation and multiple-reference editing. [FLUX.2 release][flux2] |

The sibling [FLUX.2 download guide](/thandi_fischer/flux2-ai-powerful-image-generation-model-unveiled-1mib) covers that alternative’s setup choices. Compare them with your own source images and preservation criteria.

## What else should you know about Kontext dev open weights?

### What does FLUX.1 Kontext dev generate?

FLUX.1 Kontext dev produces edited images from an input image and text instructions. Its published 12B parameter count describes the image transformer. [Model card][card]

### Can I run FLUX.1 Kontext dev locally?

BFL publishes FLUX.1 Kontext dev weights and names Diffusers, ComfyUI, and its reference code as supported implementations. Obtain model access and choose an installation path that matches your environment. [Model card][card], [Hugging Face access][hf]

### Is FLUX.1 Kontext dev licensed under Apache 2.0?

FLUX.1 Kontext dev weights use BFL’s FLUX.1 dev non-commercial license. The license has separate provisions for the model and generated outputs. [License][license]

### Does every Kontext dev edit require LoRA training?

FLUX.1 Kontext dev supports character, object, and style references without fine-tuning, according to its model card. Test an image and instruction before preparing a separate training workflow. [Model card][card]

## Sources

- [Official Kontext dev model card][card]
- [Kontext dev open-weight release][launch]
- [Official FLUX.1 repository and access overview][repo]
- [FLUX.1 dev model license][license]
- [Diffusers FLUX and Kontext pipeline documentation][diffusers]
- [Kontext technical report by Black Forest Labs][report]
- [BFL Kontext editing documentation][editing]
- [Kontext dev Hugging Face access and model card][hf]
- [ComfyUI native Kontext dev workflow][comfy]
- [FLUX.2 release and dev capabilities][flux2]

[card]: https://github.com/black-forest-labs/flux/blob/main/model_cards/FLUX.1-kontext-dev.md
[launch]: https://bfl.ai/blog/flux-1-kontext-dev
[repo]: https://github.com/black-forest-labs/flux
[license]: https://github.com/black-forest-labs/flux/blob/main/model_licenses/LICENSE-FLUX1-dev
[diffusers]: https://huggingface.co/docs/diffusers/api/pipelines/flux
[report]: https://arxiv.org/abs/2506.15742
[editing]: https://docs.bfl.ai/kontext/kontext_image_editing
[hf]: https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev
[comfy]: https://docs.comfy.org/tutorials/flux/flux-1-kontext-dev
[flux2]: https://bfl.ai/blog/flux-2

<!-- pz-related-guides -->
## Related guides on PromptZone

- [Realistic Photos with FLUX](/stabletom/realistic-photos-with-flux-57aa)
- [Best SDXL Models in 2026](/tara_suzuki/best-sdxl-models-in-2026-realistic-anime-and-all-purpose-checkpoints-116)
- [AI Model Releases Timeline](/ai-model-releases)