17h ago

Canva launches MRT, a 20-billion parameter diffusion model that generates editable image layers instead of flat pixels

It delivers 10x to 100x faster inference speeds.

——0——
Original post
OPEthan#1884Ethan@TORCHCOMPILED

Super exciting release and really proud of all who worked on it!

5:55 PM · May 27, 2026 View on X
Reposted by
Alexia Jolicoeur-Martineau#486@JM_ALEXIA
POSTAK#28AK@_AKHALIQ

MRT

Masked Region Transformer for Layered Image Generation and Editing at Scale

3:06 PM · May 28, 2026 · 3.5K Views
REPLYAK#28AK@_AKHALIQ

paper: https://huggingface.co/papers/2605.27235

AKAK@_akhaliq

MRT Masked Region Transformer for Layered Image Generation and Editing at Scale

3:06 PM · May 28, 2026 · 3.5K Views
3:06 PM · May 28, 2026 · 2.9K Views