21 papers this week in Generative media12 active threadsbusiest: Alignment Training Dynamics

Generative media research

Robotics & embodied AI

Robots can improve actions by changing randomness used for video prediction

Liu et al. show that pretrained robot policies can perform tasks more reliably by changing the random inputs from which they generate predictions, without updating the policies themselves. Across ten model-task pairs, their video-noise steering method raised average success by 26.2 percentage points over the base policies, compared with 7.1 points for an existing action-noise steering method.

3 days ago
Generative media

Single-step image restoration matches generative quality without iterative sampling

The authors propose FDIR, a framework that trains a single-pass image restoration model using a directly computable Fréchet loss, avoiding the complexity of diffusion distillation or adversarial training. By matching feature distributions between restored and real images at both image and patch levels, and introducing instance-adaptive statistics, FDIR achieves perceptual quality comparable to multi-step diffusion methods while training only low-rank adapter parameters.

4 days ago
Generative media

Fixing a training mismatch yields better video generation faster

The authors propose BlockAlign, a framework that aligns trajectory-distilled video generators (PDD, PTD) with human preferences by constructing states at block boundaries and supervising the fused block-level velocity. This avoids a mismatch that caused naive alignment to degrade quality below baselines. On two models, BlockAlign achieves significant human preference gains, up to 82.7% win rate against naive alignment, and enables faster online training.

4 days ago