FLUX.2 [dev] Turbo LoRA
FLUX.2 [dev] Turbo is a distilled LoRA adapter for FLUX.2 [dev] that enables high-quality image generation in just 8 inference steps.
Key Features
- ⚡ 8-step inference — 6x faster than the base model's typical 50 steps
- 🎨 Quality preserved — Matches or surpasses the original FLUX.2 [dev] quality
- 🔌 LoRA adapter — Lightweight, easy to integrate with existing FLUX.2 workflows
- 🖼️ Multiple modes — Supports both text-to-image and image editing
Hosted API Endpoints
FLUX.2 [dev] Turbo is available through fal.ai hosted endpoints:
ComfyUI
ComfyUI-compatible weights are available in the comfy/ directory, converted by ByteZSzn.
Usage
import torch
from diffusers import Flux2Pipeline
# Pre-shifted custom sigmas for 8-step turbo inference
TURBO_SIGMAS = [1.0, 0.6509, 0.4374, 0.2932, 0.1893, 0.1108, 0.0495, 0.00031]
pipe = Flux2Pipeline.from_pretrained(
"black-forest-labs/FLUX.2-dev",
torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights(
"fal/FLUX.2-dev-Turbo",
weight_name="flux.2-turbo-lora.safetensors"
)
prompt = "Industrial product shot of a chrome turbocharger with glowing hot exhaust manifold, engraved text 'FLUX.2 [dev] Turbo by fal' on the compressor housing and 'fal' on the turbine wheel, gradient heat glow from orange to electric blue , studio lighting with dramatic shadows, shallow depth of field, engineering blueprint pattern in background."
image = pipe(
prompt=prompt,
sigmas=TURBO_SIGMAS,
guidance_scale=2.5,
height=1024,
width=1024,
num_inference_steps=8,
).images[0]
image.save("output.png")
License
This model inherits the FLUX [dev] Non-Commercial License from the base model.