Guidance is becoming a cheat code for LLMs (with Loooop)
Why token-level guidance differs from guidance over full sequences, and how similar predictions across loops can reduce this normalization mismatch.
Blog
Scaling Video Generation for Reasoning: At What Cost?
A controlled 2 × 2 × 2 Rubik's Cube study of hidden-state reasoning in video models, with interactive results on training compute, state accuracy, and symbolic guidance.
Compress to Remember: Learning Compact Memory via On-Policy Distillation for Long Video Generation
PACC learns compact memory for a frozen video generator through on-policy distillation. The project page presents the training method, memory-based generation, and side-by-side video comparisons.
Struggling Toward Generative Model Evaluation
A blog post on why generative model evaluation is hard, how representation-based metrics became the default language, and why benchmarks are useful but never final.
EMA: A Quiet Hyperparameter That Moves Diffusion Leaderboards
We benchmark EMA decay across pixel-, latent-, and representation-space diffusion models and show that the popular 0.9999 default silently trades recall for precision.
Taming Outlier Tokens in Diffusion Transformers
Dual-Stage Registers address outlier tokens in both the encoder and diffusion transformer, improving ImageNet-256 FID from 5.89 to 4.58 at 80 epochs.
Uni-Instruct: One-step Diffusion through Unified Divergence Instruction
A unified f-divergence framework connecting more than 10 one-step diffusion distillation methods, with a one-step FID of 1.02 on ImageNet 64 × 64. NeurIPS 2025.