acihandemir·Aug 19WAN NASIL VİDEO ÜRETİYOR?Önceki bölümlerde difüzyon modellerinin nasıl görsel ürettiğinden bahsetniştim.Basitçe akış aşağıdaki gibi ilerliyordu:
Jun Nishimura·Mar 16Pico-MoEDiT: How Class-Routed MoE Gave a Tiny DiT a 43% Quality Boost for Emoji Generation“One change cut our loss by 43% — and it wasn’t scaling up the model, the data, or the compute. It was giving each emoji category its own…
Jun Nishimura·Mar 11Why Stable Diffusion 3 Switched to Rectified Flow: A Visual ExplorerBuilding a tiny Diffusion Transformer from scratch in JAX to see why Stable Diffusion 3, Flux, and Sora all switched to Rectified Flow.
Jun Nishimura·Feb 16The Architecture Behind Every State-of-the-Art Image Generator: DiT + Rectified Flow, ExplainedThe two ideas behind Stable Diffusion 3, Flux, and Sora — how they work, and what makes them tick.
ComfyAI·Jan 12NeurIPS25 oral DiT training speed up new SOTA: REGA recent collaboration between Nankai University and Harvard University has produced a standout NeurIPS 2025 Oral paper that might just…
