FLUX.2 Klein Capability Map¶

Reference for what FLUX.2 Klein 9B can do natively, via official LoRAs, via fal.ai LoRAs, and via community extensions. Updated April 2026.

Model Variants Quick Reference¶

License: Klein 9B = Non-Commercial. Klein 4B = Apache 2.0.

Operations Map¶

Official BFL LoRAs¶

Available from Black Forest Labs, non-commercial (same as base model):

fal.ai Commercial LoRAs¶

Available via fal.ai API (commercial licensing):

fal.ai API Usage¶

import fal_client

# Text-to-image
result = fal_client.subscribe("fal-ai/flux-2/base", {
    "prompt": "...",
    "image_size": "landscape_4_3",
    "num_inference_steps": 4,  # distilled
    "guidance_scale": 1.0,
})

# Image editing (reference conditioning)
result = fal_client.subscribe("fal-ai/flux-2/edit", {
    "prompt": "change background to winter forest",
    "reference_image_url": "https://...",  # source image
    "num_inference_steps": 20,
})

# Move LoRA
result = fal_client.subscribe("fal-ai/flux-2-move", {
    "image_url": "https://...",  # image with colored bbox painted on
    "prompt": "Move the vase to the right side of the table",
    "lora_scale": 1.25,
    "num_inference_steps": 8,
})

Community LoRA Categories¶

Anatomy/Quality Control¶

Face/Portrait¶

Style¶

Hundreds of community style LoRAs. Discovery via: - CivitAI (Klein 9B filter) - huggingface.co (search "flux2-klein") - Standard loading: ComfyUI LoRA loader, strength 0.4-0.75

Multi-LoRA Stacking Rules¶

Max concurrent LoRAs: 3 (practical limit before artifacts)
Total combined strength: <1.5-2.0
Stacking order: strongest-influence first

Example safe stack:
  Character LoRA: 0.6
  Anatomy Slider: 0.8 (effective, scale 2.0 not strength)
  Detail Enhancement: 0.5
  Total influence: moderate, compatible

Inference Endpoints Comparison¶

Gotchas¶

• 4B LoRAs incompatible with 9B: different hidden dims, text encoders. Always verify model variant before loading LoRA.
• Anatomy slider scale 2-4, NOT 0-1: klein_slider_anatomy v1.5 uses a wider scale range than typical LoRAs. Using standard 0-1 range produces no visible correction.
• BFL LoRAs require base model for training: if you need to fine-tune on top of an official LoRA effect, you must train on the base model — the LoRA itself is not a fine-tuneable checkpoint.
• KV variant (9B-kv) for multi-reference only: the KV cache variant provides 2.5× speedup specifically when conditioning on reference images. For standard text-to-image, there is no speedup.
• fal.ai endpoint availability: BFL periodically updates or renames endpoints. Check fal.ai/models for current endpoint IDs before hardcoding.

See Also¶

• flux klein 9b architecture - block structure, Qwen3 encoder, 3D RoPE
• flux klein 9b inference - sampler settings, CFG values, multi-pass upscaling
• flux klein character lora - character/identity LoRA training
• diffusion lora training - general LoRA training pipeline
• object removal inpainting - EraDiff, TurboFill, LanPaint