FLUX.1 Kontext is a new suite of generative flow matching models for in-context image generation and editing, developed by Black Forest Labs (BFL). It expands upon traditional text-to-image models by unifying instant text-based image editing with generation, allowing multimodal prompting with both text and images. This enables seamless extraction and modification of visual concepts for new, coherent renderings.

The core methodology revolves around multimodal flow models, leveraging a "diffusion transformer" architecture, as evidenced by the FLUX.1 Kontext [dev] variant. Unlike previous generative models primarily focused on text-to-image synthesis, FLUX.1 Kontext is designed to understand and create from existing images, making it a multimodal generative system.

Specifically, the model learns a continuous transformation or vector field that maps a simple prior distribution (e.g., Gaussian noise) to complex image data, or from a source image to a target edited image, conditioned on multimodal inputs. In flow matching, this often involves learning a velocity field $\mathbf{v}_t(\mathbf{x}_t, \mathbf{c})$ that describes the optimal path from $\mathbf{x}_0 \sim p_0(\mathbf{x})$ to $\mathbf{x}_1 \sim p_1(\mathbf{x} | \mathbf{c})$, where $\mathbf{c}$ represents the conditioning information (text prompts, reference images, or a combination). The image generation or transformation can then be obtained by numerically solving an ordinary differential equation (ODE) $\frac{d\mathbf{x}}{dt} = \mathbf{v}_t(\mathbf{x}_t, \mathbf{c})$. The "diffusion transformer" implies that this velocity field is parameterized by a transformer architecture, which excels at processing sequential or spatial data and integrating diverse contextual information through attention mechanisms.

Key technical capabilities include:

1. Multimodal Conditioning: The models accept both text prompts ($\mathbf{t}$) and reference images ($\mathbf{I}_{ref}$) as input. This implies a joint embedding space or cross-modal attention mechanisms $\text{Attention}(\mathbf{Q}, \mathbf{K}, \mathbf{V})$ within the transformer architecture to integrate $\mathbf{t}$ and $\mathbf{I}_{ref}$ into the conditioning $\mathbf{c}$.
2. In-context Generation and Editing: The ability to modify an input image ($\mathbf{I}_{in}$) via text instructions suggests that the model learns to steer the flow from $\mathbf{I}_{in}$ towards a target image $\mathbf{I}_{out}$ that satisfies the text prompt, while maintaining fidelity to unmentioned aspects of $\mathbf{I}_{in}$. This can be formulated as $\mathbf{I}_{out} = \text{Flow}(\mathbf{I}_{in}, \mathbf{t}, \mathbf{c}_{visual})$, where $\mathbf{c}_{visual}$ might represent learned visual concepts from $\mathbf{I}_{in}$.
3. Character Consistency: This implies the model can encode and preserve specific visual identities (e.g., facial features, object shapes) across different scenes or transformations. This might involve robust identity embedding or attention mechanisms focused on specific character tokens/regions.
4. Local Editing: The model can make targeted modifications without affecting other parts of an image. This suggests spatial conditioning or masking capabilities, where the model's output velocity field $\mathbf{v}_t$ is constrained or guided to specific image regions, potentially through region-of-interest (ROI) masking or attention applied to specific pixel coordinates or semantic segments.
5. Style Reference: Generating new scenes while preserving unique styles from a reference image, directed by text prompts, suggests the model can disentangle style information from content and apply it to new compositions. This is often achieved through style encoders or adaptive instance normalization (AdaIN) layers within the generative network, conditioned by the style embedding from $\mathbf{I}_{ref}$.
6. Iterative Editing: The model allows for sequential modifications while preserving quality and consistency. This indicates the flow process is robust to re-initialization with previous outputs, maintaining coherence and character identity over multiple generation/editing turns. The fast inference speed (up to 8x faster than leading models like GPT-Image, and an order of magnitude faster for FLUX.1 Kontext [pro]) is crucial for this iterative workflow, possibly due to optimized sampling schedules or the inherent efficiency of flow matching compared to multi-step diffusion sampling.

The suite includes FLUX.1 Kontext [pro] for fast, iterative editing and text-to-image generation with multimodal input, and FLUX.1 Kontext [max] for maximum performance in prompt adherence and typography. An open-weight, lightweight 12B "diffusion transformer" variant, FLUX.1 Kontext [dev], is available for research and customization. Performance is validated on KontextBench, a crowd-sourced benchmark, showing strong results in text editing and character preservation, alongside superior inference speeds.