FLUX.2 klein¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
flux2-klein-4b |
black-forest-labs/FLUX.2-klein-4B | e7b7dc2 |
Apache-2.0 | Verified |
flux2-klein-9b |
black-forest-labs/FLUX.2-klein-9B | 92196c8 |
FLUX Non-Commercial License | Verified |
Catalog: flux-2 ·
Models: flux2-klein-4b, flux2-klein-9b · Task: t2i
Architecture. The 4B checkpoint uses a Qwen3-4B text encoder, and its FLUX.2 transformer uses 5 double-stream blocks followed by 20 single-stream blocks. The 9B checkpoint uses a Qwen3-8B text encoder, and its FLUX.2 transformer uses 8 double-stream blocks followed by 24 single-stream blocks. Both text encoders supply hidden states from layers 9, 18 and 27, which are stacked for prompt conditioning. The FLUX.2 VAE has 32 latent channels; its latents are patchified 2×2 and normalized with batch-norm statistics. The flow-match Euler scheduler uses an empirical shift. The native image size is 1024×1024.
Usage.
from imagehub import pipeline
with pipeline(
"t2i", backend="flux2-klein-4b", model="checkpoints/flux2-klein-4b", device="mps"
) as pipe:
result = pipe("a cat, oil painting", seed=42)
result.images[0].save("cat.png")
Settings. steps defaults to 4, guidance defaults to 1.0 and width/height default to
1024. The other supported options are num_images, seed and latents. Unrecognized settings
raise an error.
Seed compatibility. On CPU, the same seed and settings give the same image as the pinned
diffusers reference 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. Image editing, condition images and fp16/bf16 parity are not supported.
License. The 4B weights use the Apache-2.0 license. The 9B weights use the FLUX Non-Commercial License; see the license. imagehub does not distribute weights.
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.

referenceimagehub

referenceimagehub