• Efficiency isn't squeezed out afterward by distillation. Mage-Flow co-designs the tokenizer and backbone as a pair that accommodate each other. At 4B, on a single A100, the Turbo variant renders a 1024² image in 0.59 seconds — pulling text-to-image plus instruction editing into a self-hostable, fine-tunable range.

• Give action commands in the model's own language. Masked Visual Actions writes actions as partially revealed trajectories in pixel space, dropping the external action encoder. One checkpoint does both forward dynamics prediction and inverse action inference.

• Successful and failed samples shouldn't share one supervision scheme. H²SD switches the privileged teacher's role by trajectory outcome: correct rollouts reuse their own reasoning path and only retune credit, while wrong ones bring in a verifier reference for corrective distillation.

Also Notable

• Text Prompts Can't Give Creatives the Fine Regional Control They Need — Material, object identity, spatial arrangement are all hard to pin down in one prompt. Appearance Pointers feeds multimodal regional pointers straight into the DiT to specify them.

Read the full edition →