This research introduces "Constant Acceleration Flow (CAF)" - a new way to make image generation faster and better.
CAF accelerates diffusion model generation from thousands of steps to just one. Adding acceleration to ordinary differential equation (ODE) flows makes one-step image generation possible
https://arxiv.org/abs/2411.00322
🤖 Original Problem:
Diffusion models, while excellent at image generation, suffer from slow generation speed due to their multi-step process. Existing solutions like Rectified Flow try to speed up by straightening trajectories with constant velocity, but face accuracy issues when paths intersect (flow crossing problem).
-----
🔧 Solution in this Paper:
Constant Acceleration Flow (CAF) introduces acceleration as a learnable variable alongside velocity in the ordinary differential equation (ODE) framework. It uses two key techniques:
→ Initial Velocity Conditioning (IVC): Provides acceleration model with flow direction information to handle path intersections better
→ Reflow Process: Enhances initial velocity learning through deterministic coupling
→ Closed-form solution enabling efficient one-step sampling
-----
💡 Key Insights:
→ Negative acceleration models perform better for image sampling
→ Flow characteristics can be controlled by adjusting initial velocity scale (h)
→ IVC helps overcome non-intersecting condition requirement in Rectified Flow
→ Acceleration modeling with IVC effectively addresses flow crossing problem
-----
📊 Results:
→ CIFAR-10: Achieves FID score of 1.39 (conditional) and 1.48 (unconditional) with one sampling step
→ ImageNet 64x64: Records FID score of 1.69 with one step
→ Shows 3.37 PSNR improvement over Rectified Flow in coupling preservation
