Thoughts on generative models, video quality, and the intersection of perception and computation.
Best-of-N search pays full price for every candidate it throws away. This post shows that the candidate rankings search depends on survive aggressive training-free caching, so you can explore every candidate cheaply and re-generate only the winner at full compute: 94.7% of best-of-8's gain at 63% of its cost, with any cache engine and any verifier.
Read more →What three video-generation projects on TACC's Vista and Lonestar6 taught us about GPU infrastructure: memory hierarchy and OOM taxonomy, calibration-first batching, queue-driven multi-node fleets, asynchronous RL orchestration, telemetry that tells the truth, and the measured utilization evidence behind each lesson.
Read more →A long revision notebook for Research Scientist / GenAI loops focused on image generation, perceptual quality, and video processing: diffusion and flow models, transformer internals (attention variants, RoPE, KV cache), the text-to-image design space, evaluation metrics, color and HDR, classical CV, RL alignment, and the coding tier, all in one place.
Read more →Why classifier-free guidance breaks on flow models, and how a geometry-aware predictor-corrector fixes it, with a deep dive into text rendering quality and the adaptive schedule that makes it work.
Read more →An intuition-first deep dive into Rectified Flow with the core equations, geometric interpretation of trajectory straightness, and minimal PyTorch-style code snippets for training and sampling.
Read more →Why measuring quality in high dynamic range videos is fundamentally different from SDR, and what makes it one of the most challenging problems in perceptual quality research.
Read more →