Recent Articles

• Jun 2, 2026

  Training a Pixel-Space DiT in 26 Hours: FP8 Breakthroughs and Architectural Dead Ends

  Following our integration of Asymmetric Flow Matching, our 400M parameter NanoDiT was training efficiently in terms of step-count convergence, but it was hitting an iter_per_sec of 0.025 on our single...
• May 23, 2026

  prx-tg: Accelerating Pixel-Space Diffusion with Asymmetric Flow Matching

  The evolution of text-to-image synthesis is currently undergoing a profound architectural realignment. For years, the dominant paradigm relied on Variational Autoencoders (VAEs) to compress high-dimensional pixel data into a mathematically...
• May 13, 2026

  Training a Portrait DiT on a Single GPU: What the Ablation Study Taught Us

  The prevailing assumption in generative AI is that training a large, multi-modal Diffusion Transformer from scratch requires a cluster. prx-tg is a direct challenge to that assumption: a 400M+ parameter...
• May 5, 2026

  Eradicating Syntax: Building a Neural Universal Machine That Executes Graphs, Not Code

  Our GNN autoencoders achieved 81% node accuracy on Ruby ASTs yet produced 0% valid code. The culprit was the literal value bottleneck — nearly half of every AST consisted of...
• Apr 20, 2026

  The Literal Value Bottleneck: Why GNN Autoencoders Fail at Code Generation

  We trained GNN autoencoders on 22,000 Ruby ASTs. The models achieved 81% node type accuracy and 99.5% type diversity, yet generated exactly 0% syntactically valid code. Here is why.
• Apr 20, 2026

  Self-Supervised Pretraining Recipes for Lung CT: A Systematic Study with DINO

  Self-supervised learning (SSL) promises to unlock the diagnostic potential of large unlabeled medical image archives, yet practitioners face a daunting hyperparameter landscape with little domain-specific guidance. We present a systematic...