landmark
FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space
Nikolai Kalischek, Michael Gharbi, Frederic Boesel · Black Forest Labs · 2025-06 · arXiv:2506.15742
Why it matters
Handles generation and editing in one rectified-flow model by the simplest possible mechanism: concatenate the reference image's tokens onto the generation sequence and let attention do the rest. No adapters, no inversion, no task-specific heads. Became the open reference point for in-context editing.
What this paper does
read: full textBefore this
Local editing tools such as LaMa, latent-diffusion inpainting, RePaint, Stable Diffusion inpainting, and FLUX.1 Fill relied on explicit masks. Instruction-based generative editors such as IP-Adapter, InstructPix2Pix, Emu Edit, OmniGen, HiDream-E1, and ICEdit were trained on synthetic edit pairs and inherited whatever flaws those synthetic generation pipelines had. Autoregressive multimodal editors folded into systems like GPT-Image and Gemini's native image generation gave lower quality than denoising approaches and ran too slowly for interactive use. Across all of these, keeping a character's or object's appearance stable across repeated edits remained an open, largely unsolved problem.
The problem
Unify generation and editing in a single flow-matching model, handling both local edits and open-ended in-context generation tasks, without inversion, without a task-specific architecture per capability, and without the identity drift that accumulates across multi-turn editing.
The idea
Take the reference image's latent tokens and simply append them to the sequence being generated, then let the model's existing attention layers work out the correspondence between reference and target. No adapter module, no inversion procedure, and no dedicated editing head are introduced.
How it works
The edit happens in VAE latent space; FLUX.1 is a rectified-flow transformer operating on a 16-channel latent autoencoder. Context image tokens are appended to the target image's tokens and processed jointly, first through double-stream blocks that keep separate weights per modality, then through 38 single-stream blocks where full self-attention runs over the whole concatenated sequence. 3D RoPE gives target tokens position (0,h,w) and context tokens position (i,h,w) for i=1..N, a constant offset that separates context from target as a "virtual time step" while leaving each image's own 2D spatial layout intact. There is no explicit mask, region-specific loss, or other locality mechanism. Preservation of un-edited regions is not architecturally enforced; it is a learned behavior of joint full self-attention over the concatenated sequence alone. Training starts from a pretrained FLUX.1 text-to-image checkpoint and jointly fine-tunes on image-to-image and text-to-image tasks under the rectified-flow objective; the pro variant is further trained with LADD, and the dev variant is guidance-distilled into a 12B diffusion transformer.
Evidence
The paper introduces KontextBench, 1026 image-prompt pairs across five tasks (416 local editing, 262 global editing, 92 text editing, 63 style reference, 193 character reference). Using AuraFace embedding cosine similarity, FLUX.1 Kontext [pro] scores highest on character preservation among compared systems including GPT-Image-1 and Runway Gen-4, and shows slower identity drift than those systems across repeated sequential edits. Inference reaches 3 to 5 seconds per 1024x1024 image for both text-to-image and image-to-image, up to an order of magnitude faster than the compared systems. A separate Internal-T2I-Bench of 1,000 prompts scores prompt following, aesthetics, realism, typography accuracy, and speed.
Limitations
The authors concede that excessive multi-turn editing introduces visible artifacts, showing degraded output after six iterative edits. They note the model "occasionally fails to follow instructions accurately," that distillation "can introduce visual artifacts that impact the fidelity of the output," and give concrete failure cases such as modifying a scene globally (adding milk foam) instead of performing the requested local edit (repositioning a mug). They list extending to multiple image inputs, further scaling, reducing inference latency for real-time use, and reducing multi-turn degradation as open problems.
Why it matters
Sequence concatenation plus ordinary full attention, with no adapters, no inversion, and no task-specific heads, was enough to match or beat specialized editing systems on character consistency and speed. This made the concatenate-and-attend mechanism the open reference point that subsequent in-context editing work is measured against.
Abstract, in the authors' own words
We present evaluation results for FLUX.1 Kontext, a generative flow matching model that unifies image generation and editing. The model generates novel output views by incorporating semantic context from text and image inputs. Using a simple sequence concatenation approach, FLUX.1 Kontext handles both local editing and generative in-context tasks within a single unified architecture. Compared to current editing models that exhibit degradation in character consistency and stability across multiple turns, we observe that FLUX.1 Kontext improved preservation of objects and characters, leading to greater robustness in iterative workflows. The model achieves competitive performance with current state-of-the-art systems while delivering significantly faster generation times, enabling interactive applications and rapid prototyping workflows. To validate these improvements, we introduce KontextBench, a comprehensive benchmark with 1026 image-prompt pairs covering five task categories: local editing, global editing, character reference, style reference and text editing. Detailed evaluations show the superior performance of FLUX.1 Kontext in terms of both single-turn quality and multi-turn consistency, setting new standards for unified image processing models.
Research lines
Design-axis choices
Conditioning & control
Method note — the shared flow-matching interpolation
Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.