Inclusion AI releases IDM method: 1–2 step image generation without DINOv2

🤖 Inclusion AI releases IDM method: 1–2 step image generation without DINOv2

The Intrinsic Distribution Matching method complements DMD distillation by aligning feature moments from an EMA copy of the generator itself, eliminating the need for an external encoder like DINOv2: at inference, only the generator, text encoder, and VAE remain. The arXiv:2610.39633 paper was published on October 29, 2026, and the code is available under the Apache 2.0 license.

🌍 On Z-Image-Turbo 6B and Qwen-Image 20B with 2 NFE, the method claims 94.25% on LongText-CN and 90.20% on LongText-EN—the best results among compared methods; 0.25–0.30 seconds per image for the 6B model on a single A100 (author measurements) in BF16.

👤 Code with a ready-made inference.py is already on GitHub: an environment for generating images with text (posters, menus, covers) can be prepared in advance. However, weights for Z-Image and Qwen-Image have not yet been released, and the license for future weights is unknown. Source 1: https://arxiv.org/abs/2610.39633 Source 2: https://github.com/inclusionAI/IDM