Model Releases
Black Forest Labs Releases FLUX.1 Kontext [dev] Open-Weight In-Context Image Editing Model
Black Forest Labs launches FLUX.1 Kontext [dev], a 12B parameter model enabling high-fidelity, instruction-based image editing while maintaining scene consistency.
May 28, 2025
Model ReleaseFLUX.1 Kontext
Introduction
On May 28, 2025, Black Forest Labs shifted the paradigm of generative image editing with the release of FLUX.1 Kontext [dev]. While traditional text-to-image models excel at creating new scenes from scratch, they often struggle with the iterative nature of creative workflows. FLUX.1 Kontext addresses this gap by introducing a true in-context image generation and editing framework.
This release marks a significant milestone for the open-weight community. By providing a model that understands both text instructions and visual context, Black Forest Labs is empowering developers to build sophisticated tools for precise, multi-round image manipulation that was previously locked behind proprietary APIs.
- Release Date: May 28, 2025
- Model Type: In-context image generation and editing
- Availability: Open-weight [dev] version via Hugging Face
Key Features & Architecture
At its core, FLUX.1 Kontext [dev] is a 12B parameter model built upon a generative flow matching architecture. Unlike standard diffusion models that rely solely on text embeddings, the Kontext family is multimodal by design. It treats images as part of the prompt context, allowing the model to 'see' the existing scene before applying changes.
The architecture is optimized for high-fidelity instruction following. It enables users to perform complex edits—such as changing an object's color, adding elements, or altering lighting—while maintaining strict character and scene consistency across multiple rounds of modifications. This prevents the 'drifting' effect common in iterative prompting.
- Parameter Count: 12B
- Architecture: Generative Flow Matching
- Multimodal Input: Simultaneous text and image prompting
- Consistency: Preserves character and environmental identity across edits
Performance & Benchmarks
FLUX.1 Kontext [dev] has demonstrated superior performance in maintaining semantic alignment compared to previous iterations in the FLUX.1 family. In qualitative assessments, the model shows a marked ability to adhere to nuanced natural language instructions, such as 'make the lighting more cinematic' or 'replace the coffee cup with a glass of water,' without altering the background texture.
While specific MMLU or HumanEval scores are not applicable to image models, the model was evaluated using a custom instruction-following benchmark for visual editing. This setup measured the 'Edit Fidelity Score' (EFS) and 'Context Preservation Rate' (CPR). The model achieved industry-leading scores in maintaining the spatial layout of the original image while successfully integrating new visual tokens.