inclusionAI Open-Sources IDM: 2-Step Distillation Beats 8-Step Z-Image-Turbo on Chinese Text
bdsqlsz · x · 2026-11-01
inclusionAI released IDM (Intrinsic Distribution Matching), an improved DMD distillation that compresses text-to-image models to 1–2 steps by moment-matching the generator's own EMA features. On Z-Image, 2-step generation renders long Chinese text better than the 8-step Z-Image-Turbo. Apache-2.0 code is open on GitHub (with a paper), weights coming soon; built with Westlake University, Tsinghua and others.
More from Multimodal
- Beginner tutorial: MiniMax H3 + Pika Flex LoRA generates stereo-audio video in 60s — No-Sleep-4069 · 2026-11-01
- Pika launches H3 Flex: faster-than-real-time video model, up to 80% cheaper than H3Max — techNmak · 2026-11-01
- DiffusionART-Flash open-sourced: RL for diffusion models with 25x fewer training timesteps — YangYou1991 · 2026-11-01
- Redditor shares full storyboard prompt for a 3-minute robot short film made with Sonata 5.5 — Ford-X · 2026-11-01
- Reddit's viral AI short 'Gnomes' Balloon' charms with a gnome and glowing jack-o'-lantern — love1008 · 2026-11-01
- Shizuku AI unveils in-house singing lip sync tech, beating uLipSync in quality — bdsqlsz · 2026-11-01