WaiT: Wavelet-Aware Flow Matching Sets New SOTA in Pixel-Space FID
chaumian · x · 2026-08-04
The paper "WaiT for the Signal" introduces WaiT, a frequency-aware flow-matching architecture designed to solve the conflict between global coherence and local detail in high-resolution image generation.
Core Innovations:
- Wavelet Decomposition: Decomposes generation into coarse and fine bands via lossless wavelets.
- Delayed High-Frequency Generation: High-frequency details stay pure noise until coarse structures emerge, then join the flow for joint refinement.
Evaluation Performance:
- Introduces a stringent native-resolution three-axis evaluation protocol to fix standard FID's loss of fine detail via downsampling.
- Achieves a pixel-space FID of 1.43 on ImageNet 512x512, is Pareto-optimal across all axes, and reduces sampling compute by up to 50%.
- Its largest 2B model sets a new state-of-the-art FID of 1.3 for pixel-space models on ImageNet 512. The architecture also scales seamlessly to video generation, achieving an SOTA FVD of 0.84 on Kinetics-600.
More from Multimodal
- A ComfyUI workflow pushes Wan 2.2 image-to-video out to roughly 45 seconds — embryo10 · 2026-08-04
- A tiny H3 demo detail: the weights shift during the conversation — Oatilis · 2026-08-04
- AI-made “Ancient China” video draws attention for its stylized visual design — xiaosun86 · 2026-08-04
- Qwen Image Edit crashes in ComfyUI when blur nodes are added first — UnorthodoxyMedia · 2026-08-04
- MiniMax H3 paired with LTX 2.3 for video upscaling on a 4080 GPU — alisitskii · 2026-08-04
- MiniMax H3 opens its weights as a multimodal video model with native stereo audio — mark_k · 2026-08-04