FLUX.2 [dev]

The 32B, guidance-distilled FLUX.2 model. A new double/single-stream transformer conditioned on Mistral Small 3.1 (24B) hidden states, working in the 32-channel latent space of the new FLUX.2 VAE. Reference images go in as extra tokens, so one model generates and edits.

weights: black-forest-labs/FLUX.2-dev ↗
org: Black Forest Labs
modality: image
tasks: text-to-image · image-editing · multi-reference editing
license: FLUX Non-Commercial License
released: 2025-11-25not verified
native output: 1024×1024, up to 4 MPnot verified
total params: 56.32B
model.arch: flux2

Components

role	model	params	size	dtype	trained
Transformer	FLUX.2 [dev] DiT black-forest-labs/FLUX.2-dev/flux2-dev.safetensors ai-toolkit loads this single file in the original BFL layout into its own port of the reference model (Flux2 in flux2/src/model.py). The repo also has a Diffusers transformer/ folder (Flux2Transformer2DModel) with the same parameter count.	32.22B 32,223,281,152	64.45 GB	bf16	yes
Text encoder	Mistral Small 3.1 24B Instructnot verified black-forest-labs/FLUX.2-dev/text_encoder transformers.Mistral3ForConditionalGeneration The full vision-language model, vision tower included. ai-toolkit loads mistralai/Mistral-Small-3.1-24B-Instruct-2503 instead, whose sharded files have the same parameter count and byte size.	24.01B 24,011,361,280	48.02 GB	bf16	no
Tokenizer	Mistral / Pixtral processor (131k vocab) black-forest-labs/FLUX.2-dev/tokenizer transformers.PixtralProcessor ai-toolkit uses the processor from the Mistral repo, with fix_mistral_regex off.	—	—	—	—
VAE	FLUX.2 VAE (32 channels) black-forest-labs/FLUX.2-dev/ae.safetensors Not the FLUX.1 VAE. Original-layout file; the int64 part is the BatchNorm step counter. Also shipped as a Diffusers vae/ folder (AutoencoderKLFlux2). ai-toolkit loads ae.safetensors from name_or_path.	84.0M 84,046,372	336 MB	fp32+int64	no
total		56.32B	112.81 GB

Latent space

spatial

8×

channels

patch

2×2

autoencoder: FLUX.2 VAE
pixels per token: 16×16
notes: The 2×2 patchify is part of the VAE (patch_size [2, 2] in its config), followed by BatchNorm with stored running statistics instead of a fixed scale and shift. ai-toolkit's VAE wrapper does both inside encode, so cached latents are 128 channels at 1/16 resolution and the transformer runs with patch size 1 (in_channels 128). Each token covers 16×16 pixels, as in FLUX.1.

input	latent (c×t×h×w)	tokens
512×512	32×1×64×64	1,024
1024×1024	32×1×128×128	4,096	ai-toolkit sample default
2048×2048	32×1×256×256	16,384	4 MP

Architecture

Blocks: 8 double-stream + 48 single-stream
Hidden size: 6144 (48 heads × 128)
FFN size: 18432 (mlp_ratio 3)
Text conditioning: Mistral hidden states from layers 10, 20 and 30, concatenated to 15360-d, 512 tokens. No pooled vector
Image conditioning: Reference images VAE-encoded and appended as tokens, each offset on the RoPE time axis
Guidance: Guidance-distilled, guidance embedding
Objective: Rectified flow, resolution-dependent shift
Position: 4-axis RoPE (32, 32, 32, 32), theta 2000

In AI Toolkit

model.arch: flux2
UI label: FLUX.2 (image)
model.name_or_path: black-forest-labs/FLUX.2-dev
source: extensions_built_in/diffusion_models/flux2/flux2_model.py
extensions_built_in/diffusion_models/flux2/src/model.py
extensions_built_in/diffusion_models/flux2/src/pipeline.py
extensions_built_in/diffusion_models/flux2/src/sampling.py
toolkit/models/v2/vae/flux2_kl.py
toolkit/models/v2/text_encoders/mistral3.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: datasets.multi_control_paths, sample.multi_ctrl_imgs, model.low_vram, model.layer_offloading, model.qie.match_target_res

UI defaults

quantize / quantize_te: true / true
qtype: qfloat8
low_vram: true
train.unload_text_encoder: false
noise_scheduler / sampler: flowmatch / flowmatch
timestep_type: weighted
model_kwargs.match_target_res: false
network.conv: disabled (linear LoRA only)

Specifics

Gated: FLUX.2-dev is gated: accept the license on the Hub and set HF_TOKEN. The Mistral repo is not gated.
Transformer source: flux2-dev.safetensors from name_or_path (a Hub repo or a local folder containing that file), loaded into ai-toolkit's own Flux2 module, not Diffusers.
Text encoder source: Always mistralai/Mistral-Small-3.1-24B-Instruct-2503, whatever name_or_path is. Prompts are wrapped in a chat template with a fixed system message and padded to 512 tokens. Never trained.
VAE source: ae.safetensors from name_or_path, or model.vae_path (a local file or a repo/file.safetensors Hub path).
Resolution: Buckets snap to multiples of 16.
Reference images: Up to three control images per sample (ctrl_img_1–3); datasets.multi_control_paths for training. Each is capped at 1024×1024 pixels (1 MP) with its aspect ratio kept, or resized to the target's pixel count when match_target_res is on. Only the target tokens are kept from the prediction.
Guidance during training: The guidance embedding gets train.cfg_scale (default 1.0).
low_vram: Transformer and text encoder are loaded to the CPU and moved to the GPU when needed.
Quantization: Transformer uses qtype, the text encoder uses qtype_te. layer_offloading is supported.
LoRA target: Flux2 (double_blocks and single_blocks)
Saving: LoRA keys use the diffusion_model. prefix and the original BFL module names, which ComfyUI loads. Full fine-tunes save one safetensors file in the original layout.
Metadata base version: flux2

Example config

not verified

job: extensionconfig:  name: "my_flux2_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 32        linear_alpha: 32      save:        dtype: float16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          cache_latents_to_disk: true          resolution: [512, 768, 1024]      train:        batch_size: 1        steps: 2000        gradient_checkpointing: true        noise_scheduler: "flowmatch"        timestep_type: "weighted"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16        cache_text_embeddings: true      model:        name_or_path: "black-forest-labs/FLUX.2-dev"        arch: "flux2"        quantize: true        quantize_te: true        qtype: "qfloat8"        low_vram: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        guidance_scale: 4        sample_steps: 25        prompts:          - "a bear building a log cabin in the snow covered mountains"