InclusionAI's IDM adds moment constraints to DMD, hits SOTA long-text rendering in 2 steps
inclusionAI · hf · 2026-10-30
InclusionAI, with Westlake University, Tsinghua and others, published "Intrinsic Distribution Matching" (IDM), an improvement to distribution matching distillation for few-step text-to-image generation.
- Method: Decomposes the reverse-KL field into a Gaussian moment component and a non-Gaussian residual, then explicitly constrains identifiable statistics. Feature means and per-coordinate standard deviations are matched via the generator's EMA copy — no external feature encoder and no separate fake-score network needed.
- Results: Full-parameter and LoRA tuning on Z-Image Turbo 6B and Qwen-Image 20B; with only 2 network evaluations, IDM achieves state-of-the-art long-text rendering while cutting inference cost.
More from Multimodal
- Open source image model grumble: Moss underdelivers while Krea 2 wins by default — Neggy5 · 2026-10-30
- DataTalksClub and Nebius launch free course on production AI agents for generative media — Al_Grigor · 2026-10-30
- AI-generated Hindi podcast pits Bhagat Singh against Gandhi in a virtual studio — purifiedwater1234 · 2026-10-30
- FontBox LoRA brings bbox-based intuitive text editing to FLUX.3 Klein — SillyLilithh · 2026-10-30
- Pika launches H3 Flex video model: 15s 768p clips in 7 seconds, open weights soon — heyshrutimishra · 2026-10-30
- AI-designed 'Dumpling Bag' goes viral: silicone shell nailed the pleated look — Delahuntagram · 2026-10-30