Notes by Lex Toumbourou

My WIP summary of paper FLUX.1 Kontext: Flow Matching for In‑Context Image Generation and Editing in Latent Space by Black Forest Labs

Overview

FLUX.1: Kontext is Black Forest Labs’ new text‑to‑image and image‑to‑image model, focused on image editing.

To create Kontext, they start from a FLUX.1 base checkpoint, make a few small architectural tweaks and fine‑tune with edit‑focused data.

They also introduce KontextBench, 1 026 image‑prompt pairs across five tasks (local, global, text, style, character edits) designed to test both single‑turn quality and multi‑turn stability.

FLUX 1 Kontext

Objective

Learn $p_\theta(x \mid y, c)$

• x – target image
• y – optional reference image
• c – text instruction

When y is empty it's simply text‑to‑image objective, so it can support text-to-image and image-to-image in the same model.

Key tricks

• Sequence concat – Encode x and y with the frozen FLUX auto‑encoder, then append context tokens after target tokens (channel‑wise concat was worse). Works for arbitrary resolutions and multiple context images.
• 3‑D RoPE offset – Give all target tokens position $(0,h,w)$ and each context image $y_i$ a constant “virtual‑time” offset $(i,h,w)$. This cleanly separates context from target with no extra masking.
• Rectified‑flow loss – $L_{\theta} = \mathbb{E}_{t,x,y,c} \left\| v_{\theta}(z_t, t, y, c) - (\varepsilon - x) \right\|_2^2$ where $z_t = (1 - t)x + t\varepsilon, \quad \varepsilon \sim \mathcal{N}(0, 1)$ and $t$ follows a resolution-shifted logit-normal schedule. Dropping y tokens recovers the standard T2I loss.
• Fast sampling via LADD – Latent Adversarial Diffusion Distillation halves the step count while improving quality and avoids over‑saturated CFG artifacts.

Training setup

Start from a FLUX 1 T2I checkpoint, fine‑tune on millions of ⟨context, image, prompt⟩ triplets (dataset * Three variants:

• pro – full flow objective → distilled with LADD
• dev – 12 B diffusion transformer distilled from pro, image‑to‑image only
• max – like pro but with extra compute

Safety & infra

• CLIP‑style classifier filtering plus adversarial training against NCII/CSEM.
• FSDP‑2 mixed‑precision (bf16 gather / fp32 reduce), selective activation checkpointing, Flash‑Attention 3, region‑wise compilation.

Results

Text‑to‑image

On KontextBench, Recraft v3 tops Aesthetics and Realism. GPT‑Image (high) rules Prompt Following and Typography, matching my notes in Imagen 4 is faster, but GPT is still the GOAT.

FLUX.1 Kontext is a near‑top contender in every metric despite not being tuned purely for T2I.

Image‑to‑image editing

Kontext dominates all five editing tasks, especially multi‑turn consistency.

See my own test: Absurdly Good Doggo Consistency with FLUX.1 Kontext.