ArcFlow: 2-Step Image Generation via Non-Linear Flow Distillation

🎯 The Problem

Diffusion models need many steps (20-50+) to generate quality images. Existing distillation methods use linear shortcuts between timesteps, but diffusion trajectories have constantly changing velocities that linear paths can't match.

💡 Key Insight

ArcFlow uses non-linear trajectories parameterized as a mixture of continuous momentum processes. This captures velocity evolution and enables analytical integration — no discretization errors!

⚡ The Result

Built on Qwen-Image-20B and FLUX.1-dev, ArcFlow achieves 40× speedup with only 2 NFEs, fine-tuning <5% of parameters. Quality is preserved with no significant degradation.

📊 Trajectory Visualization

0 Noise

1 Step 1

2 Image

Teacher (50 steps)

Linear Distillation

ArcFlow (Non-linear)

🔬 Why Non-Linear Works

Linear Shortcut Error

0.00

ArcFlow Error

0.00

Velocity Field Modeling

ArcFlow parameterizes the velocity field as a mixture of continuous momentum processes:

v(x, t) = Σᵢ αᵢ(t) · mᵢ(t)

Where each momentum mᵢ evolves based on learned dynamics. This enables the trajectory to match the teacher's changing velocities at each timestep.

Velocity Match

Path Length

0.00

Curvature

0.00

Analytical Integration

Unlike numerical methods with discretization errors, ArcFlow admits closed-form integration of the non-linear trajectory. This means high-precision approximation of the teacher's multi-step path in just 2 steps.

📈 Benchmark Results

Method	NFEs	GenEval ↑	DPG-Bench ↑	Speedup
FLUX.1-dev (Teacher)	50	0.67	84.2	1×
Linear Distillation	2	0.52	71.8	25×
Consistency Distillation	2	0.58	76.4	25×
ArcFlow (Ours)	2	0.65	82.9	40×