Junoh Kang

Back

Slides PDF Open slides ↗

Overview#

This post reviews three papers related to test-time scaling which enable optimization with restpect to diverse condtions without fine-tuning diffusion models nor training external models.

  • Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-Based Decoding
  • Test-Time Alignment of Diffusion Models without Reward Over-Optimization
  • Inference-Time Scaling for Flow Models via Stochastic Generation and Rollover Budget Forcing
Search