IDM Distillation Hits Best 2-NFE Long-Text Rendering: 94.25% CN / 90.20% EN
_akhaliq · x · 2026-10-30
The Intrinsic Distribution Matching (IDM) paper improves few-step text-to-image distribution matching distillation (DMD) by adding an explicit moment loss on the generator's own EMA feature means and per-coordinate standard deviations—no external feature encoder needed. The motivation comes from decomposing the reverse-KL field into a Gaussian moment component and a non-Gaussian residual.
- At 2 NFE, IDM (Z-Image) scores 94.25% on LongText-CN and 90.20% on LongText-EN, best among compared methods at this budget
- Validated with full-parameter and LoRA tuning on Z-Image-Turbo 6B and Qwen-Image 20B
- Ablations show consistent gains from the moment loss; EMA features are competitive with external DINOv2
- Code open-sourced at github.com/inclusionAI/IDM; model weights coming soon, from inclusionAI
Related event: IDM Distillation Achieves SOTA Long-Text Rendering in Two Steps(2 posts)→
More from Multimodal
- ElevenLabs Recast swaps any character in your footage while everything else stays identical — mhdfaran · 2026-10-30
- AI-generated 1972 soul performance 'Self Esteem' is convincingly retro — Mother-Pear-2392 · 2026-10-30
- Three AI-animated shorts in, character consistency is still the hardest part (MiniMax H3 / LTX / ComfyUI) — magratheya_64 · 2026-10-30
- Image captioning with zero image-text pairs: one rotation maps DINO features into text space — randall_balestr · 2026-10-30
- The Keeper: a 20-second AI-generated farewell to a giant robot — Opening-Ad5541 · 2026-10-30
- YuE2 music model impresses: one-shot tracks in Beatles, MJ, Queen styles — VasaFromParadise · 2026-10-30