Skip to content

FLUX Kontext

Image editing model from Black Forest Labs (BFL). Extends FLUX.1 architecture with sequence concatenation for context image conditioning. Best-in-class for text editing, character consistency, and multi-turn editing.

Paper: arXiv:2506.15742 (June 2025).

Architecture

Same 12B DiT as FLUX.1-dev: - 19 double-stream blocks (separate image/text weights, joint attention) - 38 single-stream blocks (fused FFN) - Hidden: 3072, heads: 24, 16-channel latent VAE - Text: CLIP (768d) + T5 (4096d)

Key Innovation: Sequence Concatenation

Context images encoded by FLUX VAE → latent tokens appended to target image tokens in visual stream: 

Target tokens:  position (0, h, w)
Context image:  position (1, h, w)   ← "virtual timestep" via 3D RoPE
Context image 2: position (2, h, w)  ← architecturally supported, not yet released

Channel concatenation was tested and performed worse. Sequence concat preserves independent resolution/aspect ratio for input vs output.

When context is empty → falls back to pure text-to-image.

vs [[Step1X-Edit]]

Aspect FLUX Kontext Step1X-Edit / Qwen-Edit
Base FLUX.1 12B DiT Custom MMDiT
Text encoder CLIP + T5 Qwen 2.5 VL (vision-language)
Conditioning Sequence concat (latent tokens) Joint attention (image + text)
Mask support No explicit mask (text-driven) DiffSynth pipeline supports masks
Speed 3-5s at 1024×1024 Slower (~40 GB VRAM)

Variants

Variant Access License
Kontext [dev] Open weights (HF) Non-commercial (BFL dev license)
Kontext [pro] API only Commercial via BFL
Kontext [max] API only Commercial via BFL

Dev is I2I only (no T2I). T2I only in pro/max.

Training

  • Starts from FLUX.1 T2I checkpoint
  • Joint fine-tune on I2I + T2I with rectified flow loss
  • Data: millions of relational pairs (not disclosed)
  • FSDP2, Flash Attention 3, selective activation checkpointing

VRAM

~24 GB minimum (transformer in bf16). Full pipeline ~32-40 GB. FP8/FP4 quantization available for 24 GB GPUs.

Gotchas

  • Modifies entire image, not just edited region. Use flux-kontext-diff-merge node for selective merging via LAB color space detection + Poisson blending
  • Multi-turn editing degrades after ~6 iterations
  • Dev model has distillation artifacts
  • Non-commercial license for open weights — commercial requires BFL licensing
  • [[ACE++]] development suspended on FLUX base due to training instability
  • GitHub: github.com/black-forest-labs/flux
  • HF: huggingface.co/black-forest-labs/FLUX.1-Kontext-dev
  • Diff-merge: github.com/safzanpirani/flux-kontext-diff-merge