FLUX.1
Esta página aún no está disponible en tu idioma.
FLUX.1 is Black Forest Labs’ family of rectified-flow transformer text-to-image models. It uses two text encoders (T5-XXL and CLIP-L) and a 16-channel VAE. InvokeAI detects three variants on install:
- FLUX.1 dev — guidance-distilled. Uses a Guidance value rather than classifier-free guidance. FLUX.1 Krea dev and FLUX.1 Kontext dev are dev-architecture models and are handled the same way.
- FLUX.1 schnell — timestep-distilled for 4 steps. It ignores guidance.
- FLUX.1 Fill dev — an inpainting/outpainting model (shown as FLUX Dev - Fill). It cannot do text-to-image.
The family is supported by a set of add-on models: ControlNets, Control LoRAs, an IP-Adapter, FLUX Redux and PiD super-resolution decoders.
License
Section titled “License”| Model | License | Commercial use |
|---|---|---|
| FLUX.1 schnell | Apache 2.0 | Allowed |
| FLUX.1 dev, Kontext dev, Krea dev, Fill, Redux and the Canny and Depth Control LoRAs | FLUX.1 [dev] Non-Commercial License | The weights may not be used commercially. Images you generate may be, except to train a competing model. |
| T5 and CLIP text encoders, SigLIP image encoder | Apache 2.0 / MIT | Allowed |
The starter models download from ungated copies where possible. The FLUX Fill, Redux and Kontext models, and the control LoRAs, come from Black Forest Labs’ own gated repositories: accept the license on each model’s HuggingFace page and add your HuggingFace token under API Keys in the Model Manager.
This is a summary, not legal advice. Read the full license on the model’s HuggingFace page before using a model commercially.
Hardware
Section titled “Hardware”See the System Requirements table. FLUX.1 needs about 10 GB of VRAM and 32 GB of system RAM at 1024×1024. Download totals with dependencies are roughly 12 GB for the NF4 and GGUF Q4 builds, 18 GB for the dev int8 and Krea dev fp8 builds, 20 GB for the Krea dev int8 build and 33 GB for bfloat16. The dev int8 transformer occupies about 11.5 GB on any supported GPU, so it streams on a 10 GB card.
To fit a full-precision checkpoint in less memory, see FP8 Storage and Low-VRAM mode.
Installing
Section titled “Installing”The simplest path is the FLUX.1 bundle in the Model Manager. It installs:
- FLUX.1 schnell (quantized) and FLUX.1 dev (quantized)
- FLUX.1 Kontext dev (quantized) and FLUX.1 Krea dev (quantized)
- FLUX Fill and FLUX Redux
- the FLUX VAE, the int8 T5 encoder and the CLIP-L encoder
- the InstantX Union ControlNet, the Canny and Depth Control LoRAs, and the XLabs IP-Adapter
Only FLUX.1 schnell, the VAE and the encoders (T5, CLIP, and the SigLIP image encoder FLUX Redux uses) in the bundle allow commercial use. Its other models are FLUX.1 [dev] models or add-ons built on it, under non-commercial terms. Some download from gated repositories; see License.
Individual starter models for the main model:
| Starter model | Format | Notes |
|---|---|---|
| FLUX.1 dev / FLUX.1 schnell | bfloat16 single file | ~33 GB with dependencies (bf16 T5) |
| FLUX.1 dev (quantized) / schnell (quantized) | bitsandbytes NF4 | ~12 GB with dependencies (int8 T5) |
| FLUX.1 dev (int8) | int8 single file (community repack) | ~18 GB with dependencies; half the memory of bf16 on every supported GPU |
| FLUX.1 dev (fp8) / schnell (fp8) | Kijai’s fp8 (e4m3fn) casts, without scales | ~22 GB with dependencies (bf16 T5); the transformer is ~11.9 GB; FP8 Storage is switched on at install |
| FLUX.1 schnell (SDNQ uint4 + SVD) | SDNQ Diffusers pipeline | ~15 GB, self-contained |
| FLUX.1 Kontext dev (quantized) | GGUF Q4_K_M | ~12 GB with dependencies |
| FLUX.1 Kontext dev (fp8) | Comfy-Org scaled fp8 single file | ~22 GB with dependencies (bf16 T5); the transformer is ~11.9 GB; FP8 Storage is switched on at install |
| FLUX.1 Krea dev / Krea dev (quantized) | single file / GGUF Q4_K_M | ~29 GB / ~12 GB with dependencies |
| FLUX.1 Krea dev (fp8) | Comfy-Org scaled fp8 single file | ~18 GB with dependencies; FP8 Storage is switched on at install, about 11.9 GB once loaded |
| FLUX.1 Krea dev (int8) | int8 single file (community repack) | ~20 GB with dependencies; stays int8 on every supported GPU, about 14.1 GB once loaded |
| FLUX Fill | single file | Inpainting model |
Except for the self-contained SDNQ pipeline, a FLUX.1 main model is only the transformer, and needs three components. Installing a starter model installs them as dependencies:
| Component | Starter options |
|---|---|
| VAE | FLUX.1-schnell_ae (works with every FLUX.1 variant) |
| T5 encoder | bfloat16 (~9.5 GB), bitsandbytes int8 (~5 GB), GGUF Q6_K (~3.9 GB, near-lossless), GGUF Q3_K_S (~2.1 GB, lower quality) |
| CLIP embed | clip-vit-large-patch14 (~250 MB) |
You can also install the T5 encoder ComfyUI uses as a single file (t5xxl_fp16, or t5xxl_fp8_e4m3fn_scaled for
about half the VRAM) and select it here; see
Model Format Support. A single-file T5 run on the CPU is
held in float32 (about 19 GB of RAM for T5-XXL).
Select them in the Components section next to the model. The SDNQ pipeline carries its own encoders and VAE, so its component pickers stay empty.
Some Black Forest Labs repositories are gated on HuggingFace. If a download asks for an access token, accept the license on the model’s HuggingFace page and add your HuggingFace token under API Keys in the Model Manager.
Generation settings
Section titled “Generation settings”Selecting a FLUX.1 model applies these defaults:
| Variant | Steps | Guidance | Scheduler | Size |
|---|---|---|---|---|
| dev (incl. Krea, Kontext) | 28 | 3.5 | Euler | 1024×1024 |
| schnell | 4 | — (ignored) | Euler | 1024×1024 |
| Fill | 50 | 30 | Euler | 1024×1024 |
- Guidance — the slider next to Steps sets FLUX’s distilled guidance, not CFG. Higher values follow the prompt more closely and give less varied images.
- Scheduler — Euler, Heun (2nd order; better quality at about twice the time per step) or LCM (for few steps).
- Size — width and height must be multiples of 16.
- DyPE — Dynamic Position Extrapolation for sizes above the 1024px training resolution: Auto turns it on above 1536px, Area by image area, 4K optimized is a fixed tuning, and Manual exposes the magnitude (λs, 0–8) and decay speed (λt; about 2 for 4K, 1 around 2K–3K, 0.5 just above native). Off by default; recorded in metadata and restored by Recall All.
- Negative prompt — FLUX.1 has no negative prompt in the Generate tab or on the Canvas.
Text-to-image, image-to-image, inpainting and outpainting all work with dev and schnell models on the Canvas. Regional guidance is supported, and a regional layer can also carry a FLUX Redux reference image.
Reference images
Section titled “Reference images”A FLUX.1 model accepts up to five reference images. Three kinds are available:
- FLUX Redux — image variation from a reference. Pick a FLUX Redux model on the reference image card and set Image influence from Lowest to Highest (default). Redux needs the SigLIP image encoder, which installs with it.
- IP-Adapter — the Standard Reference (XLabs FLUX IP-Adapter v2) starter model references an image more loosely. It uses the CLIP ViT-L image encoder, installed as a dependency.
- Kontext — when the selected model is FLUX.1 Kontext dev, reference images are passed to Kontext for instruction-style editing: describe the change you want in the prompt.
ControlNets and Control LoRAs
Section titled “ControlNets and Control LoRAs”On the Canvas, a control layer can use either kind of FLUX.1 structural control:
- ControlNet — the FLUX.1-dev-Controlnet-Union starter model (InstantX) covers canny, tile, depth, blur, pose, gray and low-quality inputs. XLabs FLUX ControlNets are also recognized.
- Control LoRA — Hard Edge Detection (Canny) and Depth Map are Black Forest Labs’ official Control LoRAs. A Control LoRA cannot be combined with a FLUX Fill model.
FLUX Fill
Section titled “FLUX Fill”FLUX Fill regenerates a masked region of an image. The Generate tab and Canvas refuse a Fill model (they
report that it does not support text-to-image), so use it in the workflow editor: connect a
FLUX Fill Conditioning node (image plus inpainting mask) to the Fill Conditioning input of
FLUX Denoise, and connect the VAE to its controlnet_vae input. A guidance of about 30 is
recommended.
FLUX.1 LoRAs can be applied from the Generate tab or with the Apply LoRA - FLUX and Apply LoRA
Collection - FLUX nodes. They patch the transformer and, where the LoRA includes text-encoder layers, the
CLIP and T5 encoders. Supported layouts include LyCORIS, Diffusers/PEFT, OMI and ComfyUI/OneTrainer PEFT files
whose module names start with diffusion_model.. Per-layer .alpha values in these diffusion_model. files set
the LoRA scale.
PiD super-resolution decode
Section titled “PiD super-resolution decode”FLUX.1 latents can be decoded with a PiD decoder, which produces a 4× super-resolved image. Install PiD Decoder FLUX (2K), PiD Decoder FLUX (2K to 4K) or the newer PiD 1.5 Decoder FLUX (2K to 4K) (also as an int8 build); see PiD Super-Resolution Decode.
Workflow nodes
Section titled “Workflow nodes”| Node | Purpose |
|---|---|
| Main Model - FLUX | Loads the transformer, T5, CLIP and VAE |
| Prompt - FLUX | Encodes the prompt |
| FLUX Denoise | Runs the diffusion; has inputs for ControlNet, Control LoRA, IP-Adapter, Redux, Kontext and Fill conditioning |
| Latents to Image - FLUX / Image to Latents - FLUX | VAE decode / encode |
| Kontext Conditioning - FLUX | Wraps a reference image for Kontext |
| FLUX Kontext Image Prep | Resizes the first image to a preferred Kontext resolution and joins up to ten images side by side |
| FLUX Redux, FLUX IP-Adapter, FLUX ControlNet, Control LoRA - FLUX, FLUX Fill Conditioning | Conditioning for the features above |
In workflows, FLUX Denoise also exposes CFG Scale with a negative conditioning input (CFG is off at the default of 1.0), and the same DyPE preset, scale and exponent as the Generate panel.