FLUX.1 Kontext [dev]

The instruction-editing member of FLUX.1. Same 12B transformer shape, text encoders and VAE as FLUX.1 [dev]; the input image goes in as a second set of tokens next to the image being generated. Trained in ai-toolkit on before/after pairs.

weights: black-forest-labs/FLUX.1-Kontext-dev ↗
org: Black Forest Labs
modality: image
tasks: image-editing · text-to-image
license: FLUX.1 [dev] Non-Commercial License
released: 2025-06-26not verified
native output: 1024×1024, output size follows the input imagenot verified
total params: 16.87B
model.arch: flux_kontext

Components

role	model	params	size	dtype	trained
Transformer	FLUX.1 Kontext [dev] DiT black-forest-labs/FLUX.1-Kontext-dev/transformer diffusers.FluxTransformer2DModel Same parameter count and config as FLUX.1 [dev] (in_channels 64, guidance embeds). Also shipped as flux1-kontext-dev.safetensors in the original layout.	11.90B 11,901,408,320	23.80 GB	bf16	yes
Text encoder	CLIP ViT-L/14 text model black-forest-labs/FLUX.1-Kontext-dev/text_encoder transformers.CLIPTextModel Only the pooled output (768-d) is used.	123.1M 123,060,480	246 MB	bf16	no
Tokenizer	CLIP BPE (49k vocab) black-forest-labs/FLUX.1-Kontext-dev/tokenizer transformers.CLIPTokenizer	—	—	—	—
Text encoder 2	T5 v1.1 XXL (encoder only) black-forest-labs/FLUX.1-Kontext-dev/text_encoder_2 transformers.T5EncoderModel 512 tokens in ai-toolkit.	4.76B 4,762,310,656	9.52 GB	bf16	no
Tokenizer 2	T5 SentencePiece (32k vocab) black-forest-labs/FLUX.1-Kontext-dev/tokenizer_2 transformers.T5TokenizerFast	—	—	—	—
VAE	FLUX.1 VAE (16 channels) black-forest-labs/FLUX.1-Kontext-dev/vae diffusers.AutoencoderKL Encodes both the target and the input (control) image.	83.8M 83,819,683	168 MB	bf16	no
total		16.87B	33.74 GB

Latent space

spatial

8×

channels

patch

2×2

autoencoder: FLUX.1 VAE
pixels per token: 16×16
notes: Same packing as FLUX.1. The control image is encoded the same way and appended as extra tokens, so with one control image at the target size the sequence is twice the token counts below.

input	latent (c×t×h×w)	tokens
512×512	16×1×64×64	1,024	target only; ×2 with control
1024×1024	16×1×128×128	4,096	target only; ×2 with control
1344×768	16×1×96×168	4,032	landscape, target only

Architecture

Blocks: 19 double-stream + 38 single-stream
Hidden size: 3072 (24 heads × 128)
FFN size: 12288
Text conditioning: T5 hidden states (4096-d) joined into the sequence; CLIP pooled vector (768-d) added to the timestep embedding
Image conditioning: Input image latents appended as tokens; their RoPE ids are marked with 1 on the first axis
Guidance: Guidance-distilled, guidance embedding
Objective: Rectified flow, resolution-dependent shift (0.5 at 256 tokens to 1.15 at 4096)
Norm / position: QK RMSNorm, 3-axis RoPE (16, 56, 56)

In AI Toolkit

model.arch: flux_kontext
UI label: FLUX.1-Kontext-dev (instruction)
model.name_or_path: black-forest-labs/FLUX.1-Kontext-dev
source: extensions_built_in/diffusion_models/flux_kontext/flux_kontext.py
toolkit/models/v2/diffusion_models/flux.py
toolkit/models/v2/text_encoders/t5.py
toolkit/models/v2/text_encoders/clip.py
toolkit/models/v2/vae/autoencoder_kl.py
toolkit/models/flux.py
toolkit/train_tools.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: datasets.control_path, sample.ctrl_img

UI defaults

quantize / quantize_te: true / true (qfloat8)
noise_scheduler / sampler: flowmatch / flowmatch
timestep_type: weighted
network.conv: disabled (linear LoRA only)

Specifics

Gated: Accept the license on the Hub and set HF_TOKEN before the first run.
Training data: Pairs: datasets.control_path holds the input images, with the same file names as the edited targets in folder_path. Captions are the edit instructions.not verified
Control image: Resized (bilinear) to the target crop size, VAE-encoded, then packed and appended after the target tokens. Only the target tokens are kept from the prediction, so the loss covers only them.
Resolution: Buckets snap to multiples of 16.
Loading: Transformer from name_or_path. T5, CLIP and VAE come from extras_name_or_path, which defaults to name_or_path, or from name_or_path if it is a local folder with text_encoder/.
Quantization: Transformer uses qtype, T5 uses qtype_te. CLIP is never quantized. layer_offloading is supported.
low_vram: Keeps the text encoders on the CPU and moves them to the GPU only to encode prompts.
Guidance during training: The guidance embedding gets train.cfg_scale (default 1.0).
LoRA target: FluxTransformer2DModel
Saving: LoRAs use Diffusers/PEFT keys (transformer.…lora_A / lora_B), alpha set to the rank. Full fine-tunes save the transformer/ folder plus aitk_meta.yaml.
Sampling: Every sample prompt needs --ctrl_img. The control image is resized to the sample width × height, which are floored to multiples of 16.
Metadata base version: flux.1_kontext

Example config

not verified

job: extensionconfig:  name: "my_first_flux_kontext_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 16        linear_alpha: 16      save:        dtype: float16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/edited/images"          control_path: "/path/to/input/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          cache_latents_to_disk: true          resolution: [512, 768]      train:        batch_size: 1        steps: 3000        gradient_checkpointing: true        noise_scheduler: "flowmatch"        timestep_type: "weighted"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16      model:        name_or_path: "black-forest-labs/FLUX.1-Kontext-dev"        arch: "flux_kontext"        quantize: true        quantize_te: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        guidance_scale: 4        sample_steps: 20        prompts:          - "make the person smile --ctrl_img /path/to/input/images/person1.jpg"