New Paper: Contrastive Noise Alignment Cuts FID Over 50% in Few-Step Generation
burny_tech · x · 2026-09-18
A new arXiv paper introduces Contrastive Noise Alignment (CNA), a training-time method that optimizes Gaussian noise representations for diffusion and flow-matching models.
- Problem: Independently sampled noise induces arbitrary data-noise couplings, forcing high-curvature transports; existing OT methods only reassign noise samples, leaving the noise distribution passive.
- Method: Models the noise batch as an interacting particle system, using a cross-modal InfoNCE objective to align noise particles with paired data, regularized by angular entropy and radial norm penalties.
- Theory: The equilibrium asymptotically preserves Gaussian structure, keeping inference tractable.
- Results: For 2-4 NFE pixel-space generation, CNA reduces FID by over 50% vs standard rectified flow and at least 24% vs OT baselines, with no added inference cost.
Authors: Lennart Wittke and Vinicius Azevedo; 21 pages.
More from Multimodal
- Claude picked the Voynich Manuscript, failed to crack it, then made a surprisingly good explainer video — emollick · 2026-09-18
- MiniMax-H3 in Practice: 10s 1MP Clips in Under 6 Minutes on a 5070Ti — Sleepy_Bandit · 2026-09-18
- ComfyUI Workflow Makes MiniMax H3 AI Video 5X Faster — cointalkz · 2026-09-18
- PrunaAI launches P-Video-2-Pro: 5s 480p video in ~2 seconds, from $0.02/s — Kyrannio · 2026-09-18
- New Pika launches as all-in-one AI creative platform that handles model plumbing — Kyrannio · 2026-09-18
- 15-second short 'Talos' made with Seedance 2.5 and Minimax Music 3.0 inside new Pika — Kyrannio · 2026-09-18