Long form content,my own research notes, and thoughts on vision & ML ·
A series exploring emergent capabilities hidden inside VL-conditioned single-stream diffusion transformers. No fine-tuning, no additional training — just scaffolding: architectural hacking and careful probing of what these models already know.
Vision-only token dropout solves mode collapse in spliced I2I generation. Along the way: hunting attention sinks, discovering which conditioning tokens are load-bearing, and two orthogonal knobs for controlling diversity.
Z-Image Base and Turbo are text-to-image models, but they're secretly image-to-image models too. A simple architectural splice unlocks zero-shot image variations with no training.
Lorem ipsum dolor sit amet. This template shows how to use tables and other formatting options.