The VAE is dead; long live patchify-plus-diddly. Linum throws out the separately-trained VAE entirely and folds compression into the DiT itself via aggressive 32×32 patchification (512×512 image → 256 tokens). The key enabling trick is x-prediction with v-loss: everyone's been regressing velocity, which forces the model to fit high-dimensional Gaussian noise, and that's why DiTs historically choke on high-dimensional inputs. Predict the clean image instead, weight by the velocity term, and the curse of dimensionality mostly evaporates. Their fine-detail fix is the JiT-DDT: a conditional encoder predicts an 8× downsampled 64×64 "plan" that a decoder uses, plus a mid-training widening of the noise schedule toward low-noise timesteps (2.3% of timesteps below t=0.1 in phase 2, up from under 0.05%). Net claim: 3.6× fewer GPU-hours than their Linum v2 baseline at 4× the pixels. Apache 2.0 weights, faces still malformed, gradients now run through everything. Notably, DINOv3 outperforms Qwen3.5-4B at its one job, which must sting in all the wrong ways.