Skip to content

FLUX.1

Model Source Revision Weight license Parity (CPU, fp32)
flux1-schnell black-forest-labs/FLUX.1-schnell 741f7c3 Apache-2.0 Verified
flux1-dev black-forest-labs/FLUX.1-dev 3de623f FLUX.1-dev Non-Commercial License Verified
flux1-krea-dev black-forest-labs/FLUX.1-Krea-dev 8162a9c FLUX.1-dev Non-Commercial License Verified
flux-1-lite-8b Freepik/flux.1-lite-8B 16b9373 FLUX.1-dev Non-Commercial License Verified
flux1-kontext-dev (+ image input) black-forest-labs/FLUX.1-Kontext-dev 24e9ded FLUX.1-dev Non-Commercial License Verified

Catalog: flux-1 · Models: flux1-schnell, flux1-dev, flux1-krea-dev, flux-1-lite-8b, flux1-kontext-dev · Task: t2i

Architecture. CLIP-L supplies a pooled embedding and T5-XXL supplies sequence embeddings to a 12B rectified-flow transformer with 19 double-stream blocks, 38 single-stream blocks, 24 attention heads of width 128 and 3-axis RoPE. The dev checkpoint adds a guidance embedding. A 16-channel VAE with a shift factor decodes latents packed in 2×2 patches. The flow-match Euler scheduler uses a static shift of 1.0 for schnell and dynamic shifting for dev. The native image size is 1024×1024.

Usage. Both repositories are gated. Accept the repository terms and provide HF_TOKEN when downloading:

HF_TOKEN=... python -m imagehub.pretrained flux1-schnell checkpoints/flux1-schnell
from imagehub import pipeline

with pipeline(
    "t2i", backend="flux1-schnell", model="checkpoints/flux1-schnell", device="mps"
) as pipe:
    result = pipe("a cat, oil painting", seed=42)
    result.images[0].save("cat.png")

Settings. Schnell defaults to 4 steps, guidance 0.0 and a maximum sequence length of 256. Dev defaults to 50 steps, guidance 3.5 and a maximum sequence length of 512. Both default to 1024×1024, and width/height must be multiples of 16. The other supported options are num_images, seed, latents, negative_prompt and true_cfg_scale. Unrecognized settings raise an error.

Validation. On CPU, the same seed and settings give the same image as diffusers with generator=torch.Generator("cpu").manual_seed(seed) (validation). MPS and CUDA runs are smoke tests only; there is no claim of exact equality on these devices.

Out of scope. Fill, Redux, Depth and Canny variants, ControlNet and fp16/bf16 parity are not supported. flux1-schnell, flux1-dev, flux1-krea-dev and flux-1-lite-8b take no image input; for Kontext, see below.

License. The schnell weights use the Apache-2.0 license. The dev, Krea [dev] and Kontext weights use the FLUX.1-dev Non-Commercial License; see the license. imagehub does not distribute weights.

Flux.1-Lite 8B

flux-1-lite-8b (Freepik/flux.1-lite-8B) distills FLUX.1-dev by reducing the double-stream transformer from 19 blocks to 8 while retaining all 38 single-stream blocks, the two text encoders, VAE and dynamically shifted flow-match scheduler. The component weights are published only in bfloat16 and are upcast for CPU float32 inference.

python -m imagehub.pretrained flux-1-lite-8b checkpoints/flux-1-lite-8b
from imagehub import pipeline

with pipeline(
    "t2i", backend="flux-1-lite-8b", model="checkpoints/flux-1-lite-8b", device="mps"
) as pipe:
    result = pipe("a cat, oil painting", seed=42)
    result.images[0].save("cat.png")
  • Defaults. 28 steps, guidance 3.5, 512 text tokens and 1024x1024, matching the model card.
  • Access and license. The repository is ungated. The weights inherit the FLUX.1-dev Non-Commercial License.
  • Memory. The two transformer shards total about 16.3 GB in bfloat16. CPU float32 loading requires roughly twice that for the transformer, plus T5-XXL, CLIP-L and the VAE.

FLUX.1 Krea [dev]

flux1-krea-dev (black-forest-labs/FLUX.1-Krea-dev) is an aesthetics-focused fine-tune of the FLUX.1 [dev] architecture, per its model card. Its published configs are identical to those of flux1-dev, so it runs on the same backend, text encoders, transformer and VAE with no new code.

from imagehub import pipeline

with pipeline(
    "t2i", backend="flux1-krea-dev", model="checkpoints/flux1-krea-dev", device="mps"
) as pipe:
    result = pipe("a cat, oil painting", seed=42)
    result.images[0].save("cat.png")
  • Defaults. 28 steps (the diffusers FluxPipeline default; the model card gives no step count), guidance 4.5 (the model card), max_sequence_length 512 and 1024×1024. The other options are the same as for flux1-dev.
  • License and access. The repository is gated and the weights use the FLUX.1-dev Non-Commercial License. Accept the terms and provide HF_TOKEN when downloading.
  • Memory. The same as flux1-dev: about 47.6 GB for the transformer, 19 GB for T5-XXL, 0.5 GB for CLIP-L and 0.3 GB for the VAE in float32.

FLUX.1 Kontext [dev]

flux1-kontext-dev (black-forest-labs/FLUX.1-Kontext-dev) is the FLUX.1 transformer trained to edit an image from a text instruction. It shares the FLUX.1 backend, text encoders and VAE; the only difference is that the tokens of the encoded input image are appended to the sequence, and only the generated tokens are decoded.

from PIL import Image
from imagehub import pipeline

with pipeline(
    "t2i", backend="flux1-kontext-dev", model="checkpoints/flux1-kontext-dev", device="mps"
) as pipe:
    result = pipe("Add a red hat to the scene", image=Image.open("photo.png"), seed=0)
    result.images[0].save("edited.png")
  • The image is optional. Without image, the model generates from the prompt alone, like flux1-dev.
  • Input resizing. The input is converted to RGB (grayscale, RGBA and palette images included), rounded down to a multiple of 16, and then resized to the preferred Kontext resolution whose aspect ratio is closest to its own (672×1568 up to 1568×672, 1024×1024 for a square). It is encoded by the VAE (the mode of the posterior, shifted and scaled) before it joins the sequence.
  • Output size. The size of the result comes from height and width (default 1024×1024), not from the input image. They are rescaled to at most max_area pixels (default 1024²) at the same aspect ratio and rounded down to multiples of 16. To keep the input's shape, pass its size as height and width.
  • Defaults. 28 steps, guidance 3.5, max_sequence_length 512, true_cfg_scale 1.0 (no negative prompt). The model card's examples use guidance 2.5; pass guidance=2.5 to use it. The other FLUX.1 options (num_images, seed, latents, negative_prompt, true_cfg_scale) work as for flux1-dev; max_area is the one added option besides image.
  • License and access. The repository is gated and the weights use the FLUX.1-dev Non-Commercial License. Accept the terms and provide HF_TOKEN when downloading.
  • Memory. The same as flux1-dev: about 47.6 GB for the transformer, 19 GB for T5-XXL, 0.5 GB for CLIP-L and 0.3 GB for the VAE in float32.
  • Limits. One input image only. Inpainting (FluxKontextInpaintPipeline), several input images, IP-Adapter and fp16/bf16 parity are not supported.
  • Seed compatibility. On CPU, the same seed, image and settings give the same image as diffusers 0.40.0 FluxKontextPipeline (validation).

Reference and imagehub, side by side

Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.